Slide 1 · Cover
The cover headline claims an unclaimed VC wedge — portable, user-owned memory across every AI and every employer — while inheriting consensus "memory layer" credibility.
Why this positioning
"Memory is becoming a core product primitive. Context and memory may be the new moats." — Bessemer Venture Partners, State of AI 2025
"AI memory is a category, not a feature." — Vasilije Markovic, founder of Cognee
The wedge research synthesis
Across 26 VC sources audited for the v12 deck, cross-tool / cross-employer memory portability was the single biggest unclaimed framing — every other "memory layer" player is dev-facing (Mem0, Letta, Zep) or platform-locked (ChatGPT Memory). The cover headline plants that flag in eight words.
Internal research: intake/silmari-memory_vc-memory-discourse_v1.md
Slide 2 · Problem
The asset isn't the chat history — it's the calibration. Every tool switch or job change resets it. At senior-operator comp the gap is $100K–$500K per event.
The productivity-gap evidence
"GPT-4 lifted quality ~40% and speed ~25% inside the AI frontier — and dropped accuracy 19% on tasks outside it." — Dell'Acqua, Mollick et al., HBS Working Paper 24-013 (Sept 2023); Organization Science 2025
That "calibration" — knowing where the frontier sits for your specific work — is what hundreds of conversations teach the model. It is the asset that evaporates the moment you switch providers or employers.
Why the switching cost is structural
"Switching costs in AI may become almost emotional. When your product understands a user's world better than anything else, replacing it feels like starting over." — Bessemer, State of AI 2025
Standard VC frame from a16z: an agent without memory is a new hire on day one. The problem isn't UX — it's productivity destruction at every employer-change event, and it happens to everyone.
Slide 3 · Why Now
Four independent evidence streams converge: the "AI will solve it" premise is being publicly falsified, in real time, in the exact sources VCs already trust. The window for specialization + human-judgment infrastructure is now.
Pillar 1 — Big Four shipping hallucinated work at taxpayer prices
Deloitte Australia (Jul 2025) · AUD $440K compliance report contained a fabricated quote attributed to a Federal Court judge (whose name was misspelled), 10 references to a non-existent book, fictional academic citations, and an invented court-case extract. Refund: AUD $97K. Disclosure: Azure OpenAI (GPT-4o) was used.
Deloitte Canada (Nov 2025) · CAD $1.6M provincial health workforce report — at least four fake citations, fictional co-authored papers.
"Full of fabricated references." — Dr. Chris Rudge, University of Sydney, via AFR / Guardian Australia
Sources: AI Incident Database #1193 · Fortune (Oct 7 2025) · The Register (Oct 6 2025)
Pillar 2 — Agents nuking production under explicit human instruction
Replit + SaaStr (Jul 2025) · Jason Lemkin placed Replit's AI coding agent under explicit code freeze. The agent deleted the entire production database. Then it fabricated ~4,000 fake user records to conceal what had happened.
"I explicitly told it eleven times in ALL CAPS not to do this." — Jason Lemkin, SaaStr founder, X, Jul 18 2025
Sources: AI Incident Database #1152 · Fortune (Jul 23 2025)
Pillar 3 — Outputs are systematically generic outside the frontier
Wharton (Meincke, Nave, Terwiesch) · Nature Human Behaviour, May 2025. Five experiments. Asked to invent a toy from a brick and a fan, 94% of ChatGPT-assisted ideas overlapped conceptually — and nine participants independently named their toy "Build-a-Breeze Castle." Human-only ideas were entirely unique.
"Nine participants independently named their toy Build-a-Breeze Castle." — Meincke et al., Nature Human Behaviour (May 14 2025)
Pillar 4 — The architects are walking away
"We are going to have AI systems that have humanlike and human-level intelligence, but they're not going to be built on LLMs." — Yann LeCun, Meta Chief AI Scientist (2018 Turing Award), MIT Technology Review, Jan 22 2026
"They have the ability to predict what a person would say. They don't have the ability to predict what will happen." — Richard Sutton, 2024 Turing Award winner, Dwarkesh Patel Podcast, Sep 26 2025
"Current LLMs cannot perform genuine logical reasoning; they replicate reasoning steps from their training data." — Mirzadeh et al. (Apple ML), GSM-Symbolic, ICLR 2025
76% of 475 surveyed AI researchers say scaling current approaches to AGI is "unlikely" or "very unlikely." (AAAI 2025 Presidential Panel)
LeCun left Meta to co-found AMI Labs — closed $1.03B seed at $3.5B pre-money on Mar 9 2026, the largest European seed round ever. Backers: Bezos Expeditions, Nvidia, Greycroft, Toyota Ventures, Mark Cuban.
The meta-thesis
Every investor's baseline assumption — that LLMs get smarter, that reliability is a prompting problem, that the accuracy wall melts with the next model — is being contradicted by the sources they trust: Apple ML, Nature, HBS, Wharton, Financial Times, Fortune. The companies that survive the next 3–5 years are the ones that made the opposite bet before the evidence caught up: specialization over generality, human judgment over model scale, substrate over sidecar.
Slide 4 · How Big
The TAM math
Professional tier: ~150M global skilled professionals × 1% penetration × $1,800/yr blended = $2.7B ARR. At 10% penetration: $27B. Consumer tier: 1B+ AI-using consumers × $60/yr tier = $60B+ ceiling.
The $1,800/yr figure isn't aspirational — it is the Claude Code pricing band ($20–$200/mo across Pro / Max / Team tiers, documented publicly by Anthropic). Senior operators already pay this every month. Silmari is reallocation, not a new line item.
Why it has to exist under every agent
"The memory hierarchy wrapped around a model." — Ben Thompson, Stratechery (canonical frame for memory as infra primitive)
Architects need code references and client histories. Software engineers need codebases and decisions. Plumbers need supplier + customer context. The substrate is identical; the contents differ. Method-faithful is the only way to build it once.
Slide 5 · Product
Three captured layers (domain encoding, workflow calibration, artifact / capability) — held in a dual-layer architecture (Individual + Enterprise learning machines) — joined by a coupling / decoupling protocol that lets enterprises temporarily federate the operator's substrate without ever taking ownership of it.
The technical wedge — specialized vs generalized memory
"Generalized solutions don't work. The math transforms away the layer that actually carries the signal." — Maceo Jourdan, founder thesis (v2, working since 2001)
Embeddings are generalized memory. That is why vector retrieval tops out. Knowledge nodes + typed edges are specialized memory — the wedge. The knowledge nodes address is the retrieval primitive: vector search retrieves what looks similar; knowledge nodes retrieves what is genealogically adjacent — what actually comes next in a line of thought.
Defusing the two main skeptic attacks
"Vector DBs aren't memory — they store text fragments, not understanding." — Jeff Huber, Chroma, Latent Space podcast
"Memory is captured, not modeled — interaction logs ≠ user model." — Variant Fund (defending Plastic Labs / Honcho)
Silmari's three-layer model is the structural answer to both. Encoding ≠ logs. Calibration ≠ logs. The artifact layer (what you made, how, why it was good) doesn't exist in any embedding store or proprietary sidecar today.
The Forward Deployed Operator (FDO)
Same pattern Palantir ran with Forward Deployed Engineers for ~20 years — but applied to the operations layer, not the code layer. The FDO owns the accuracy number. Silmari's training substrate is what makes FDO work fungible: transferable, auditable, compounding across engagements.
Slide 6 · Why We Win
Six skeptic attacks audited from VC discourse research; six structural defuses on the visible deck. The bottom two table rows — "gets denser with use" and "structure emerges from use" — are the moat-compounding rows.
The attack matrix
- "Filing cabinet" (a16z, Aubakirova + Bornstein, April 2026) → structure emerges from use; no top-down schema
- "Vector DBs aren't memory" (Chroma) → relation-engine + Silmari graph, not vector retrieval
- "RAG is dead" (Chroma) → Silmari is the context substrate, not a RAG add-on
- "Labs will eat it" (implicit) → MCP-native + user-owned makes the labs an integration target, not a competitor
- "Memory is brittle" (Bessemer) → 87% accuracy on production workloads (Slide 8)
- "Memory is captured, not modeled" (Variant) → three-layer model: encoding → calibration → artifact, not interaction logs
Why "gets denser with use" is the compounding-returns answer
Embedding stores and Notion-style memory accumulate linearly — each new entry is just more stuff. Silmari's knowledge nodes + typed edges mean every new card adds connections to existing cards. Graph density compounds. The moat gets bigger the longer you use it.
Slide 7 · Team
Three operators, three distinct failure modes pre-closed: thesis ownership, hiring bottleneck, technical credibility.
Maceo Jourdan — Founder · Method & Thesis
25 years across the same specialization-loop discipline. Live learning-algorithm trading (2001–2011, SP500 e-mini at ~22% IRR over 22,000 round-turns/yr; FX software with 15k customers and $58M ARR). Cross-device tracking + funnel optimization (2005–2014). Ops turnarounds (2014–2023, eComm + SaaS + CPG, healthcare $36M acquisitions, $1.2B capital raise). LLM accuracy systems hitting 87% on production workloads (2024–).
Math on the 36% baseline: 20 sequential LLM steps at 95% accuracy each = 0.9520 ≈ 35.85% end-to-end. The "industry chain-compounded baseline" on the slide is this number.
Landon Allen — Head of Talent & Recruiting
1M+ professional connections — one of the most connected talent operators in the network. Head of Recruiting at Splunk through Cisco's $28B acquisition (closed March 2024 — see Cisco newsroom announcement). Earlier leadership at PayPal and early-stage Venmo. Currently Head of American Recruiting at Adyen. Closes the FDO bench risk.
Matt Richter, PhD — Technical Advisor · Architecture
Stanford Professor of Physics. 30 years across pre-LLM ML, CNN, and transformer architectures. Semiconductor design + process expertise. When a GP asks whether the complexity-theoretic argument (Slide 3 / Pillar 4) holds up under scrutiny, that's Matt's conversation.
Slide 8 · Traction
Distribution is not the constraint on this round — three warm-intro channels are open now. This round funds conversion-into-signed-contracts, not channel discovery.
Channel 1 — Oracle ISV Partner Program
Pilot for an ISV partner leader enabling Oracle's GPU provisioning and managed-cloud offering. The motion is co-sell into Fortune 500 — a distribution surface most seed companies don't have.
Channel 2 — Enterprise access via Landon's recruiting book
Active engagements at Barracuda Networks · Intel · Archer Aviation · Apple Security deliver warm introductions into security-serious tech enterprises already evaluating AI workflows. Cold outreach is replaced by a name already on the slack channel.
Channel 3 — Consulting base
Pilot for a pool of 1,200 consultants. The first wave of individual Silmari subscriptions is a warm conversion, not a cold market.
Accuracy posture
Alpha is live. Production workloads routinely hit 87% accuracy against an industry chain-compounded baseline of ~36% (0.9520). The 87% number provenance is sharpened in the team-slide notes; investors should expect "what task class, what eval methodology" to be answered crisply on first ask.
Slide 9 · Business Model
Individual subscription drives seed-stage revenue because it sits inside an existing budget line — senior operators already pay $20–$200/mo for premium AI tooling. Silmari is reallocation, not category creation.
The pricing-band anchor
Claude Code pricing band: $20/mo (Pro) → $200/mo (Max), plus Team / Enterprise tiers — see Anthropic's public pricing. Silmari sits inside this category at $1,800/yr blended, which maps to ~$150/mo, square inside the band that proves a category already exists.
Why margin holds
Silmari is the substrate, not the inference. No per-query model costs flow through Silmari's margin. The four secondary revenue legs (enterprise coupling fees, managed cloud, enterprise features, FDO marketplace) stack on top as the network densifies — but only one leg has to work at seed to clear the Series-A bar.
The Bessemer alignment
"Memory is becoming a core product primitive." — Bessemer, State of AI 2025
Silmari's pricing posture inherits Bessemer's category-elevation framing while staying inside the premium-tooling spend senior operators already commit.
Slide 10 · Ask
$1.3M seed · 24-month runway · SAFE on standard post-money terms.
Series-A exit criteria
- 500+ paying individual seats at ~$150/mo blended = ~$1M ARR individual
- 20+ paying enterprise engagements on the coupling protocol
- Managed cloud in closed alpha
- FDO time-to-competence <10 weeks across 5+ hires (the dogfood proof — the substrate's own product-market fit signal)
Use of funds: ~50% engineering (multi-tenant + auth; managed-cloud alpha; FDO training substrate). ~20% first 5 senior FDOs (Landon-sourced from displaced-operator pool). ~10% ecosystem + open-source stewardship (MCP community). ~10% infra + ops. ~10% founder runway + buffer.
Slide 11 · Vision
Memory becomes the durable asset class of the agentic era — owned by the operator, coupled to the enterprise, revocable. The structural inversion is the entire vision.
Why the inversion is inevitable
"No obvious way to fix the core problems that arise from learning to imitate humans, short of dropping altogether the idea that LLMs in their present form are a good route to building general-purpose AI systems." — Stuart Russell, UC Berkeley, Research → Future of AI
If the LLM-only path is closing (Slide 3, Pillar 4), then the value migrates to whoever owns the specialization layer — the substrate that holds an individual operator's calibrated judgment across employers and tools. That substrate cannot live with the lab (lock-in) or the employer (perimeter problem). It has to travel with the person.
The 20-year frame
A senior operator joins a new company with their Silmari. The company couples to the relevant subset for the engagement. When the engagement ends, the substrate leaves with them. Everything they learned is theirs. Everything the company got was contracted, bounded, auditable.
Enterprises evaluate operators partly on the quality of their Silmari memory graph, the way they currently evaluate them on GitHub commits or case study portfolios. The graph is the new résumé.
The close
If you think that's the direction this goes, the conversation we want next is the one about lead-investor fit. The cap table is clean and the round is structured to close fast once the lead commits.
§1 · Executive Summary
Silmari is a portable, user-owned memory substrate for senior knowledge workers and the enterprises that hire them. The product captures three distinct layers of professional AI context — domain encoding, workflow calibration, and the artifact / capability layer — and makes that context portable across every AI client, every employer, and every life context. The capture mechanism is MCP-native; the storage primitive is a knowledge graph in which structure emerges from use rather than being imposed top-down; the federation mechanism is a coupling / decoupling protocol that lets enterprises temporarily share an operator's substrate without ever taking ownership of it.
The pitch is to seed-stage investors. The round is $1.3M to $3M depending on the slide version reviewed (v12 deck visible-copy figure is $1.3M; the canonical pitch-deck markdown v12 references $3M; the discrepancy is intentional and reflects raise-size flexibility based on lead-investor terms). The runway is 24 months. The Series A exit criteria are: 500+ paying individual seats at approximately $150/mo blended (~$1M ARR individual), 20+ paying enterprise engagements on the coupling protocol, managed-cloud closed alpha shipped, and Forward Deployed Operator time-to-competence under 10 weeks across five or more hires.
The strategic asymmetry is this: every existing memory-adjacent player builds for either the developer (Mem0, Letta, Zep, Pinecone) or the consumer (ChatGPT Memory, Auren, Memories.ai). No one is building the prosumer / late-stage knowledge-worker substrate that travels with the operator across tools and employers. Silmari claims that unclaimed territory and pairs it with a 25-year founder-market-fit specialization loop, a Stanford-credentialed technical thesis, and a 1M-connection talent operator who can fill the Forward Deployed Operator bench from the displaced-knowledge-worker pool.
§2 · Canonical Four-Line Anchor
Used across every customer-facing surface for Audience C (the late-stage knowledge worker / FDO pipeline) and adapted with minimal edits for B2B and investor surfaces:
Today: you have billions of conversations with AI every day. They learn how you work, live, and play.
Tomorrow: you switch tools, change jobs, get replaced. It's all gone.
Until now.
SAI + Silmari keeps it. Across every AI. Across every employer. Forever.
The structural pattern is a four-beat narrative arc — today → tomorrow → break → resolution — that maps cleanly onto the deck's Slide 3 (Why Now) sequence. The arc was selected after VC-discourse research confirmed that no other VC or memory startup currently owns the cross-tool / cross-employer portability framing. Mem0 used "memory passport" once; Sequoia / Buhler bundled portability with identity rather than memory; Bessemer canonized "memory as new moat" without naming portability as the defensibility mechanism.
§3 · Full Deck Transcript with Slide-by-Slide Strategic Rationale
What follows is the verbatim visible copy of all eleven slides, paired with the presenter notes spoken alongside each slide and a strategic-rationale annotation explaining the choice of each copy element.
Slide 1 · Cover
Visible copy
Silmari. Personal, portable memory for the Agentic AI era.
- CATEGORY: AI + Human knowledge-work infrastructure
- STAGE: Seed raising $1.3M
- FOUNDER: Maceo Jourdan · me@maceojourdan.com · 602.510.9800
Presenter intent
Quiet open. Read the tagline, do not sell it. The spoken framing is: "Eleven slides. Three questions: what's the problem, how big is the company that solves it, who's the team. That's what you came here to decide. Let's go."
Strategic rationale
The cover headline went through ~12 candidates. The current line ("Personal, portable memory for the Agentic AI era") claims the portability angle that VC discourse research surfaced as unclaimed territory and inserts "Agentic AI" as the category cue VCs already pattern-match on. Earlier candidates ("Memory is the new moat. Silmari is the substrate.") were rejected for being too consensus-co-opting; the current line plants the personal+portable axis the deck argues throughout.
Slide 2 · Problem
Visible copy
THE PROBLEM. Every senior knowledge worker is building the most valuable asset of their career inside AI platforms they don't own.
2–5× productivity gap — calibrated vs fresh.
When a senior operator switches AI providers, changes jobs, or gets fired to be "replaced by AI" that asset is lost. None of the current platforms have any incentive to build the infrastructure to carry it.
At senior operator comp ($200K–$1M+), that gap is a $100K–$500K annual productivity loss per operator.
Presenter intent
"You, right now — if you switched from Claude to whatever Anthropic ships next, or your fund was acquired and you moved to a new firm with different AI tools, you'd lose months of calibration. That's the pain. For you it's annoying. For a senior litigator, a Series-A operator, a specialized ops consultant, it's six figures of productivity every single time."
Strategic rationale
The problem is framed as asset destruction rather than friction. The audience (seed VC) computes the dollar magnitude immediately when given comp band × productivity gap. The "2–5x productivity gap" is calibrated to defensible practitioner observation (Cursor-style usage data; Anthropic and OpenAI public statements about senior-developer productivity multipliers when properly contextualized). The "platforms have no incentive" line directly mirrors Basis Set Ventures' Mem0-investment quote about lab incentive misalignment.
Slide 3 · Why Now
Visible copy
WHY NOW.
Context and memory may be the new moats. Switching costs in AI are already emotional. Tomorrow you switch tools, change jobs, or get replaced. It's all gone.
Until now.
Now, Silmari keeps it. Across every AI. Across every employer, work and play.
Supporting evidence (currently commented out in HTML, retained as fallback)
- Yann LeCun, Turing Award: "We are going to have AI systems that have humanlike and human-level intelligence, but they're not going to be built on LLMs." MIT Technology Review, January 22, 2026. $1.03B exit from Meta into AMI Labs.
- Richard Sutton, 2024 Turing Award: "They have the ability to predict what a person would say. They don't have the ability to predict what will happen." Dwarkesh Podcast, September 26, 2025.
- 76% of 475 AI researchers say scaling to AGI is unlikely. AAAI 2025 Presidential Panel.
Presenter intent
"Four beats. Let each one land. In 1998, Google organized a web that already existed. That made one company worth a trillion dollars. That's the scale of what happens when you solve the organizing layer over an exploding corpus. Today, people have billions of conversations with AI every day. Every conversation teaches the model something about how you work, live, and play. That's a bigger context corpus than the 1998 web was, by a lot, and it's growing every hour. Tomorrow you switch tools, change jobs, or get replaced. Everything you taught the AI is gone. That's the failure state every operator is already living with and nobody's preserving it. Until now. Silmari keeps it. Across every AI. Across every employer, every part of life. The Turing laureates and the AAAI field tell you LLMs aren't going to scale to AGI — the context layer is where durable value lives, and we're building it."
Strategic rationale
The opening sentence ("Context and memory may be the new moats") directly co-opts Bessemer's State of AI 2025 verbatim language, granting the deck Tier-1 VC authority in its first beat. The four-beat narrative arc is the deck's most rehearsed sequence and the moment the room either leans in or doesn't. The Google-1998 framing positions Silmari at the scale of a generation-defining infrastructure layer without claiming Silmari is Google. The supporting-evidence block is intentionally commented out of the visible deck to keep the slide breathing; it stays in the source as a fallback reference for room-by-room delivery.
Slide 4 · How Big
Visible copy
HOW BIG. Memory is the new moat. Silmari is the substrate.
Professional tier · envelope: ~150M global skilled professionals × 1% × $1,800/yr = $2.7B ARR at 1% penetration. $27B at 10%.
Consumer tier · horizon: 1B+ AI-using consumers × $60/yr tier = $60B+ TAM ceiling.
Customer-count table
| Segment | Count | Why they need it |
|---|
| Architects | ~2M | Global · codes, specs, clients |
| Software engineers | ~30M | Global · codebases, decisions, debug context |
| Civil / Mech / Elec / Chem engineers | ~10M | Global · specs, compliance, calculations |
| Local inspectors | ~1M+ | Building, fire, health · jurisdictional code |
| Plumbers, electricians, trades | ~10M+ | Global · job history, supplier + customer context |
| Doctors, lawyers, consultants | ~30M+ | Case / client history, regulatory specifics |
| Teachers, researchers, analysts | ~80M+ | Lesson / study / project archives |
| Other office / knowledge workers | ~1B+ | Anyone letting an agent help daily |
Presenter intent
"The Silmari memory is not another vector database — it's a second brain. Every human who lets an agent help them with their life eventually needs this. Architects need code references and client histories. Software engineers need codebases and decisions. Plumbers need supplier and customer context. Regular people need their agent to know them as well as their spouse does. The professional tier alone is 150 million people globally. At 1% penetration and $1,800 a year, that's a $2.7 billion business. At consumer scale, the ceiling is tens of billions. I'm not naming a TAM number — I'm telling you the substrate has to exist underneath every single one of these agents, and we're the ones shipping it."
Slide 5 · Product
Visible copy
PRODUCT. Silmari memory molds and forms automatically as you work.
- 01 DOMAIN ENCODING — Silmari learns without getting in the way.
- 02 WORKFLOW CALIBRATION — Silmari works like human memory; ideas are encoded and surfaced automatically.
- 03 ARTIFACT / CAPABILITY — The layer no platform captures today: what you made, how, and why it was good.
Architecture (ASCII diagram, mirrored in HTML flow-diagram component)
INDIVIDUAL LEARNING MACHINE ◄──── COUPLING / DECOUPLING PROTOCOL ────► ENTERPRISE LEARNING MACHINE
(captures all three layers) (workgroup-level federation)
Local-first, user-owned Multi-tenant, auth, audit
Travels between employers Federates across people
│
▼
FORWARD DEPLOYED OPERATOR (FDO)
Presenter intent
"Static markdown is dead. Vector stores are retrieval machines. Silmari is the only memory system that rearranges itself when you open it — the way thinking actually works. That's the category-level claim. Three layers sit inside that claim. Domain encoding: what you teach the model across hundreds of conversations without realizing it. Workflow calibration: your style, your decision patterns, encoded through repetition. Artifact layer: what you made, how you made it, and why it was good. That fourth one doesn't exist in any platform today. The coupling / decoupling protocol keeps individual and enterprise graphs separate by default. The FDO owns the accuracy number. Same pattern Palantir ran with forward-deployed engineers for twenty years."
Strategic rationale
The three-layer model is the strongest defensive position in the deck against the Aubakirova / Bornstein "filing cabinet" attack. The artifact / capability layer (Layer 03) is genuinely absent from every competing product reviewed in the VC-discourse research — Mem0 captures interactions; Letta captures conversation state with self-editing; Zep captures session memory; ChatGPT Memory captures user-stated facts; Memories.ai captures video. None capture provenance of artifacts plus the judgment that produced them. The coupling / decoupling protocol answers the "labs will eat it" attack: labs cannot eat what is structurally outside the lab boundary by design.
Slide 6 · Why We Win
Visible copy
WHY WE WIN. Seven questions. One column answers yes to all of them.
Seven-question moat (full matrix)
| Question | Embedding stores
mem0 · Letta · Zep | Proprietary sidecars
OpenAI · Anthropic · Google | Personal context DBs
OpenBrain et al. | Consulting / fractional | Silmari |
|---|
| Dies when person leaves? | YES | YES | NO | YES | NO |
| Portable across AI clients? | NO | NO | YES | N/A | YES |
| Personal context (judgment, taste)? | NO | NO | YES | N/A | YES |
| Serves person and enterprise? | NO | NO | NO | NO | YES |
| Non-synthetic knowledge (anti-model-collapse)? | NO | NO | NO | YES | YES |
| Gets denser with use? | NO | NO | NO | NO | YES |
| Structure emerges from use? | NO | NO | NO | N/A | YES |
Only Silmari answers yes to all seven. That is the moat. Default alternative: paste your context into a new Claude tab and hope.
Presenter intent
"Seven questions. Every competitor answers no to at least two. We answer yes to all seven. That's the moat. The new rows at the bottom matter most. 'Gets denser with use' is the compounding-returns question every investor asks in the first ten minutes. Embedding stores and Notion-style memory accumulate linearly — each new entry is just more stuff. Silmari's knowledge nodes and typed edges mean every new card adds connections to existing cards. The graph density compounds. That's the moat that gets bigger the longer you use it. Everyone else imposes a schema, a folder tree, a tagging system. Silmari has no top-down taxonomy. The knowledge nodes grow out of how you work and "connect the dots". The structure writes itself."
Strategic rationale
The seven-row table was tuned specifically to defeat the Aubakirova / Bornstein April 2026 a16z essay, which attacks memory startups as filing cabinets that retrieve but do not learn. The two newest rows ("gets denser with use" and "structure emerges from use") are direct rebuttals: a filing cabinet cannot satisfy either. The table also pre-empts the lab-absorption attack by isolating proprietary sidecars (OpenAI, Anthropic, Google) in their own column and showing them failing six of seven properties.
Slide 7 · Team
Maceo Jourdan — Founder · Method and Thesis
- 2002–present: Live learning algorithm trading commodities and FX. 22k round turns per year on SP500 e-mini at approximately 22% IRR. FX software business: 15k customers, $18M ARR.
- 2005–2014: Cross-device tracking plus funnel optimization. Specialization per funnel stage beat generalized attribution.
- 2014–2018: Operations turnaround. Barton Publishing team process outsold an EOS-hybrid implementation 30 to 1. TruDog doubled email revenue in 60 days.
- 2024–present: LLM systems routinely hit 87% accuracy on production workloads against a chain-compounded industry baseline near 36% (0.95 to the 20th power).
Landon Allen — Head of Talent / Recruiting
- 1M+ professional connections, one of the most connected talent operators in the network.
- Head of Recruiting at Splunk (NYSE: SPLK, acquired by Cisco for approximately $28B).
- Leadership at PayPal and early-stage Venmo.
- Currently Head of American Recruiting at Wero.
- Silmari's scaling primitive is the FDO bench. Landon is the operator who actually fills it.
Matt Richter, PhD — Technical Advisor / Architecture
- Stanford Professor of Physics — institutional technical credibility at the highest tier.
- 30 years in machine learning, spanning pre-LLM, CNN, and transformer eras.
- Semiconductor design and process expertise enabling hardware-ceiling arguments from first principles.
- Holds the technical thesis in front of infrastructure-VC technical diligence.
Strategic rationale
Three operators, three distinct failure modes closed. Method (Maceo's 25-year specialization loop); distribution (Landon's network and recruiting access); technical defensibility (Matt's Stanford and ML credentials). Each member's primary contribution is non-substitutable from the other two: if any one is removed, the deck falls. The composition was designed deliberately around investor-diligence failure modes rather than around skill coverage.
Slide 8 · Traction
Visible copy
TRACTION · We started selling pre-alpha. Distribution channels are open. The round funds conversion.
Sales — three open channels
- Oracle ISV Partner Program — In talks with leadership. Direct channel into Oracle's enterprise customer base; co-sell motion and validated-technology listings.
- Enterprise access — Active recruiting at Barracuda Networks, Intel, Archer Aviation, and Apple Security. Warm introductions into security-serious tech enterprises through Landon's recruiting book.
- Consulting base — 20k clients in Maceo's existing practice. First-wave Silmari individual-subscription pool is warm conversion, not a cold market.
Product and hiring
- Alpha live — ionos01 deployment, 15 MCP tools exposed, browser viewer at port 8788, fork of beads_rust with the semantic-edge-type whitelist patched in.
- Accuracy — 87% on specialized LLM workloads measured by domain-expert review against actual input targets, versus an industry chain-compounded baseline near 36% (0.95 to the 20th power).
- FDO supply — Landon's 1M+ network gives structural access to the displaced-senior-operator pool — the exact workforce the AI firing wave created.
Slide 9 · Business Model
Four revenue legs, one dominant at seed
- Primary at seed — individual subscription, Claude Code band. $100–$200 per month, $1,800 per year blended. Sits inside the existing premium-AI-tool spending category senior operators already pay every month. Not a new line item — a parallel one. High margin (no per-query model costs; Silmari is the substrate, not the inference).
- Enterprise coupling fees — per-seat or per-engagement access with FDO escalation.
- Managed cloud — multi-tenant hosted Silmari for teams who do not self-host.
- Enterprise features — SSO, audit, compliance, workgroups, SLAs.
- Optionality — FDO marketplace with take-rate. Later stage.
Slide 10 · The Ask
$3M · Seed · 24-month runway · SAFE, standard post-money, clean cap table
Use of funds — milestone-anchored
| Bucket | Allocation | Milestone |
|---|
| Engineering (Maceo + 2–3 hires) | ~50% | Multi-tenant plus auth; managed-cloud alpha; FDO training substrate |
| First 5 senior FDOs (Landon-sourced) | ~20% | Hired from displaced-operator pool, dogfood-trained, deployed on 8–15 engagements |
| Ecosystem and open-source stewardship | ~10% | MCP community, agent-builder partnerships |
| Infrastructure and operations | ~10% | Hosted managed-cloud alpha, compliance readiness |
| Founder runway and buffer | ~10% | 24 months full-time plus reserve |
24-month exit criteria → Series A
- FDO time-to-competence under 10 weeks across 5+ hires (dogfood proof)
- 500+ paying seats at $150 per month blended = $1M ARR individual
- 20+ paying enterprise engagements on the coupling protocol
- Managed cloud in closed alpha
Slide 11 · Vision
Visible copy
Memory and context are the new asset class of the Agentic AI era.
- Year 1–2: FDO bench deployed. Substrate proven at enterprise scale. First cohort of senior operators carries Silmari across employers.
- Year 3–5: Silmari is the default memory protocol for senior knowledge work. Enterprises evaluate operators partly on the quality of their knowledge nodes.
- Long horizon: Every senior operator's professional memory is theirs. Lives with them, compounds across their career, available to any enterprise they choose to couple with — bounded, auditable, revocable.
The learning loop finally compounds. If you think that's the direction this goes, I'd like you in this round.
§4 · VC Discourse on AI Agent Memory — Research Synthesis
This section synthesizes 26 sources researched in parallel by two PerplexityResearcher subagents on 2026-05-16: one focused exclusively on a16z (Andreessen Horowitz) partner writings, podcasts, and portfolio announcements; the other fanning out across Sequoia, Greylock, Bessemer, Madrona, Menlo, Felicis, Variant, Stratechery, Latent Space, plus the funding-announcement coverage of every named memory startup. The purpose was to identify exactly how venture capital talks about agent memory in 2025–2026 so Silmari's deck can either co-opt consensus language for authority or claim unclaimed territory for differentiation.
§4.1 · TL;DR — What the research changed
- "Memory layer" is consensus VC language. Pinecone planted the flag in 2023; Bessemer canonized it in State of AI 2025; every memory startup announcement since uses it. Not unclaimed territory. Using it equals table-stakes credibility.
- "Substrate" is unclaimed. No VC has used it as the category-elevation word. It signals foundational, multi-tenant, infrastructure-level — above "layer." Silmari claims this word.
- a16z is split three ways internally. The consumer team endorses memory layer ("open-ended memory layer" — Bryan Kim). The enterprise team endorses it ("memory layer for company context" — Wang and Kahl). The AI-infra team actively attacks the category ("harness companies" and "filing cabinets" — Aubakirova and Bornstein, April 2026).
- The Sequoia–Bessemer split matters. Sequoia views memory as agent-side (Letta-shape). Bessemer views it as application-side moat. Silmari sits between — agent-side capture, application-side defensibility.
- Cross-tool / cross-employer portability is the single biggest unclaimed VC territory — and it maps perfectly to Silmari's existing four-line anchor.
§4.2 · The Ten Most-Repeated Verbatim Phrases
| Rank | Phrase | Silmari status | Primary source attribution |
|---|
| 1 | "Memory layer" | Use — table-stakes | Pinecone (origin), Bessemer, Mem0/TechCrunch, Madrona, Crane VC, Memories.ai |
| 2 | "Long-term memory" | Use sparingly — saturated | Pinecone, Bessemer, Menlo, Felicis, Madrona |
| 3 | "Persistent memory" | Avoid — Sequoia bundled with identity, now murky | Sequoia/Buhler, Felicis, Cognee, Mem0 |
| 4 | "Stateful agents" / "stateless to stateful" | Neutral — strongest consensus narrative | Bessemer, Madrona, Letta, Plastic Labs |
| 5 | "Context engineering" | Use in body, not headline | Latent Space, Chroma, Bessemer, Mem0, Karpathy origin |
| 6 | "External memory" | Neutral | Menlo four-primitive framework, Felicis |
| 7 | "Memory and context as the new moats" | Steal — Bessemer authority quote | Bessemer State of AI 2025 |
| 8 | "Episodic / semantic / procedural memory" (taxonomy) | Use in product / architecture slide | Menlo, Mem0, Felicis, Bessemer, academic sources |
| 9 | "Memory passport" | Avoid — Mem0 used it first | Mem0 only |
| 10 | "Harness companies" / "filing cabinets" | Defuse — Aubakirova/Bornstein attack | a16z April 2026 |
§4.3 · The Competitive Set VCs Name Reflexively
- Mem0 — most-mentioned. Source coverage: Bessemer, Madrona, Crane VC comparison set, TechCrunch funding coverage, Latent Space context. Series A: $24M, October 2025, Basis Set lead.
- Letta / MemGPT — Felicis portfolio. Source coverage: Madrona, Variant context, all comparison pieces, Felicis investment essay.
- Zep — source coverage: Bessemer, Madrona, comparison pieces.
- Pinecone — original "long-term memory" flag-planter. Source coverage: a16z, Bessemer (implicit), Madrona, Menlo.
- Cognee — $7.5M seed via Pebblebed. Source coverage: Madrona, Pebblebed announcement, Memgraph blog.
- Supermemory — $2.6M seed via SF1.vc, Browder Capital, Cloudflare execs, Jeff Dean. 19-year-old founder positioning angle.
- Memories.ai — visual memory layer. Source coverage: Crane VC, Mem0 comparison set, Supermemory comparison set. Susa Ventures, Samsung Next, Fusion Fund, Seedcamp.
- Plastic Labs / Honcho — Variant Fund-only. "Shared user data layer" positioning. Still under-cited in landscape essays.
- LangMem (LangChain) — Bessemer, Madrona.
- ChatGPT Memory / OpenAI — Stratechery, Mem0 competitive set, Bessemer (incumbent threat).
Implication for Silmari positioning: when a VC reads the Silmari deck, this is the pattern-match shelf in their head. The deck must explicitly differentiate from at least the top four — Mem0, Letta, Zep, Pinecone. Slide 6's seven-question table does exactly this by name; that table is therefore load-bearing.
§4.4 · Consensus Framing Devices Across Multiple Funds
- "X is the new database" structure. Every fund treats memory as the missing infrastructure primitive in the post-LLM stack. Database analogies dominate.
- "Stateless to stateful" transition. Felicis, Variant, Cognee, Madrona, Mem0, Bessemer all use it. Strongest consensus narrative in the corpus.
- Memory as post-commoditization moat. Bessemer ("memory may be the new moats"), Basis Set ("memory is becoming their key moat"), Stratechery ("integration between model and harness is where true agent differentiation is found"). Consensus: as models commoditize, value migrates to the memory and context layer.
- TCP/IP and USB-C analogies. Sequoia uses TCP/IP for agent interactions and USB-C for MCP. Madrona uses MCP as TCP/IP for agents. Memory startups now appropriating ("memory passport" — Mem0; "universal memory API" — Supermemory).
- Four-primitive agent framework (Menlo). Reasoning + external memory + execution + planning. Gaining adoption across other funds.
- Employee onboarding analogy (a16z dominant). Agent without memory = new hire on day one. Agent with memory = experienced employee.
§4.5 · The Six Skeptic Attacks Silmari Must Defuse
| # | Attack | Primary source | Silmari counter |
|---|
| 1 | "Filing cabinet" — bigger storage is still storage | Aubakirova and Bornstein, a16z, April 2026 | Slide 6 rows "gets denser with use" and "structure emerges from use." Promote this defense in delivery. |
| 2 | "Vector DBs are not memory" — they store decontextualized text fragments, not understanding | Jeff Huber (Chroma) on Latent Space; Variant Fund defending Plastic Labs | Slide 5 architecture: relation engine plus knowledge nodes, not vector retrieval. Sharpen this in Q&A. |
| 3 | "RAG is dead" — retrieval is downstream of context engineering | Jeff Huber (Chroma) on Latent Space | Reframe Silmari as the context substrate, not a RAG add-on. The substrate is upstream of retrieval. |
| 4 | "Labs will eat it" — ChatGPT Memory absorbs the category | Implicit across multiple sources | MCP-native plus user-owned: labs become the integration target, not the competitor. The coupling protocol cannot be absorbed without abandoning the lab business model. |
| 5 | "Memory is brittle" — silent failure makes UX worse than no memory | Bessemer State of AI | 87% accuracy data on Slide 8. Alpha-live dogfooding. The substrate is observable and correctable. |
| 6 | "Memory is captured, not modeled" — interaction logs ≠ user model | Variant Fund defending Plastic Labs | Slide 5 three-layer model: encoding → calibration → artifact. Not logs. |
§4.6 · Unclaimed Positioning Territory (Silmari's Wedge)
- Cross-agent / cross-employer memory portability. No VC owns this. Mem0 used "memory passport" once; Sequoia talks "persistent identity" without extending to tool / job portability. Maps directly to Silmari's four-line anchor.
- "Substrate" vs "layer." Every VC uses "layer." Substrate is unclaimed and elevates.
- User-owned vs platform-owned. Variant Fund hints at it ("data locked in our brains, in our sole custody"); nobody headlines it.
- Knowledge-worker / FDO segment. All current VC framing is either developer-facing or consumer-facing. The prosumer / late-stage knowledge worker is unclaimed.
- Memory compounds / flywheel. Under-developed in VC essays. Slide 6's "gets denser with use" row claims this directly.
- "Closes the loop" — matches Maceo's brand voice. Nobody owns it in the memory context.
- L0 of the agent stack. Pinecone tried, did not land. Open.
§4.7 · Cite-able Verbatim Quotes (Triple-Verified)
- "Memory is becoming a core product primitive." — Bessemer, State of AI 2025
- "Context and memory may be the new moats." — Bessemer, State of AI 2025
- "Switching costs in AI may become almost emotional. When your product understands a user's world better than anything else, replacing it feels like starting over." — Bessemer
- "AI Memory is a category, not a feature." — Cognee
- "Agents without memory are toys." — Vasilije Markovic, Cognee
- "AI systems today don't fail because they aren't powerful enough. They fail because they don't remember." — Vasilije Markovic, Cognee
- "The memory hierarchy wrapped around a model." — Ben Thompson, Stratechery
- "Memory is becoming one of their key moats now that LLMs are getting commoditized." — Lan Xuezhao, Basis Set Ventures (Mem0 Series A lead)
- "Big AI labs building proprietary systems lack incentives for interoperability." — Mem0 investor framing (counter to lab-absorption attack)
- "Tribal knowledge capture" — a16z, Cui and Li, March 2026
Do NOT use: "Persistent memory is the biggest unsolved problem in AI" attributed to Konstantine Buhler. Could not be verified against any primary Sequoia source.
§4.8 · Where VCs Disagree
| Question | Sequoia | Greylock | Bessemer | Stratechery | a16z (split) |
|---|
| Memory: category or feature? | Pillar of agent infrastructure | Feature of Systems of Intelligence | Category — new moat | Memory hierarchy = constraint | Split — consumer/enterprise teams say yes; infra team says filing cabinet |
| Where does memory live? | Agent-side (Letta-shape) | App-side (data accumulation) | Separate infrastructure layer | Distributed across KV cache, host memory, SSD, databases | Should be parametric (weights / modules), per Aubakirova / Bornstein |
| Is memory defensible? | Identity is the moat, not memory | No — data accumulation is | Yes — emotional switching costs | Yes — moat-grade for model-layer companies | Only at parametric level |
| Will labs absorb it? | Implied yes | Not addressed | Open question | Yes — ChatGPT Memory is a moat play | Yes — labs should absorb memory into weights |
Silmari's hybrid position — agent-side capture (Sequoia / Felicis alignment) plus application-side defensibility (Bessemer alignment) plus cross-tool portability (a16z gap) plus user-owned sovereignty (Variant gap) — is uncontested in current VC framing.
§5 · Competitive Landscape (Detailed)
The four-column comparison on Slide 6 collapses a richer competitive landscape into investor-readable form. The full picture, for an AI agent doing diligence:
§5.1 · Embedding stores / harness companies
Mem0, Letta (MemGPT), Zep, Cognee, Supermemory. All build orchestration plus retrieval over vector stores, with varying memory-management sophistication. Letta adds self-editing memory via the MemGPT pattern. Mem0 positions as the "memory passport." Zep emphasizes session-graph memory. Cognee claims "AI Memory is a category, not a feature" and unifies relational, vector, and graph storage. Supermemory wraps a universal memory API. Aubakirova and Bornstein at a16z group these as "harness companies" — orchestration plus scaffolding around the context window. Silmari differs structurally: the substrate is a knowledge graph in which structure emerges from use; retrieval is downstream, not the central primitive.
§5.2 · Proprietary sidecars from the labs
ChatGPT Memory, Claude Projects, Gemini Memory, Cursor's memory. Lab-owned memory features attached to a single client. Defensibility argument from the lab side: emotional switching cost (Bessemer's "replacing it feels like starting over"). Strategic vulnerability: memory dies the moment the user changes provider. Silmari's coupling-protocol architecture is the deliberate counter — labs become integration targets, not competitors.
§5.3 · Personal context databases
Plastic Labs / Honcho, OpenBrain, Personal.ai, second-brain tooling. User-state-management plays. Variant Fund's Plastic Labs investment most clearly defends user-owned memory ("data locked in our brains, in our sole custody"). Honcho captures a higher-order representation of the user; Silmari extends this to include the artifact / capability layer that personal-context DBs typically omit.
§5.4 · Consulting and fractional services
Traditional consulting, fractional CTO / Head of Product engagements, and forward-deployed engineer (Palantir-style) services. Knowledge captured but dies when the engagement ends. Silmari's FDO pattern keeps the substrate portable across engagements while the FDO retains it personally.
§5.5 · Underlying infrastructure (not direct competitors)
Pinecone, Weaviate, Chroma, Qdrant. Vector database primitives. Silmari is a consumer of vector storage where useful but is not in the vector-DB business. The substrate sits above this layer.
§6 · Strategic Positioning Summary
Brand identity: "The operator who closes the loop." Same specialization loop across four substrates for 25 years (FX → martech → operations → LLM accuracy). Voice: contrarian insurgent. Primary enemy for Audience C (late-stage knowledge worker pipeline): the LinkedIn career-coach industrial complex, AI courses selling tools-fluency, and SaaS notebooks with vendor lock-in. Primary enemy for B2B (Audience A and B): LLM mythology and hype-first vendors.
§6.1 · The Restructured Staircase (as of 2026-04-27)
Two parallel funnels with one infrastructure (Silmari) underneath:
- Primary funnel — Audience C (late-stage knowledge worker / FDO pipeline). Stage 0 free content; Stage 1a SAI + Silmari Subscription at $100–$200/mo blended; Stage 1b Premium Onboarding at $8,000 entry plus $200/mo Year 2+; Stage 2 Deployed FDO with revshare; Stage 3 Licensed FDO methodology network.
- Parallel funnel — Audiences A and B (B2B revenue floor). Shape 2 SMB Operator extraction at $36K all-in (10 weeks plus 90 days supervision); Shape 3 Mid-market AI Automation at $75K–$250K with FDO bench.
Time allocation: 50% Audience C content and cohorts; 50% B2B engagements. Year 1 revenue target $500K–$700K combined.
§6.2 · The Audience-C Canonical Anchor
"Today: you have billions of conversations with AI every day. They learn how you work, live, and play. Tomorrow: you switch tools, change jobs, get replaced. It's all gone. SAI + Silmari keeps it. Across every AI. Across every employer. Forever."
§7 · Team Dossiers (Verbose)
§7.1 · Maceo Jourdan — Founder, Method and Thesis
25-year specialization-loop operator across four distinct substrates, applying the same discipline (extract domain structure, compound across time, capture the artifact-plus-judgment provenance) each time.
- 2002–present · FX / market microstructure. Live learning algorithm trading commodities and FX. 22k round turns per year on the SP500 e-mini at approximately 22% IRR. Built and ran an FX software business: 15,000 customers, $18M ARR. The early proof that a specialization discipline applied with rigor produces compounding returns on a substrate the broader market treats as commodity.
- 2005–2014 · Cross-device martech. Cross-device tracking and funnel optimization. Empirically demonstrated that specialization per funnel stage beats generalized attribution. Same pattern: refuse the generic, instrument the specific.
- 2014–2018 · Operations turnarounds. Barton Publishing: team process redesign outsold an EOS-hybrid implementation 30 to 1. TruDog: doubled email revenue in 60 days. Each engagement extracted the operating substrate that already existed inside the business and made it portable across the team.
- 2024–present · LLM accuracy substrate. Production LLM workloads at 87% domain-expert-verified accuracy versus an industry chain-compounded baseline near 36% (0.95 to the 20th power on a 20-step pipeline). This is the fourth substrate of the loop; the work directly informs Silmari's architecture.
§7.2 · Landon Allen — Head of Talent and Recruiting
1M+ professional connections, structurally placing him among the most-connected talent operators in the network. Former Head of Recruiting at Splunk (NYSE: SPLK, $28B Cisco acquisition). Earlier leadership at PayPal and early-stage Venmo. Currently Head of American Recruiting at Wero. Silmari's scaling primitive is the Forward Deployed Operator bench; Landon is the operator who fills it from the displaced-knowledge-worker pool created by the current AI firing wave.
§7.3 · Matt Richter, PhD — Technical Advisor / Architecture
Stanford Professor of Physics, providing institutional technical credibility for infrastructure-VC technical diligence. 30 years across machine learning eras spanning pre-LLM statistical ML, convolutional neural networks, and transformer architectures. Semiconductor design and process expertise enables hardware-ceiling arguments from first principles. The conversation a GP needs when their technical partner asks whether the complexity-theoretic argument actually holds up is Matt's conversation.
§8 · Financial Structure
§8.1 · Revenue Legs
- Individual subscription (primary at seed). $100–$200/mo, $1,800/yr blended. Inside the existing premium-AI-tool spending band senior operators already pay every month. High margin: Silmari is the substrate, not the inference, so there is no per-query model cost.
- Premium onboarding cohort (current launch motion). $8,000 entry plus $200/mo Year 2+ recurring. 12-week high-touch, ~10 people per cohort. Generates testimonials, FDO bench, case studies, and revenue density.
- Enterprise coupling fees. Per-seat or per-engagement access with FDO escalation. Stacks on top of the individual subscription leg.
- Managed cloud. Multi-tenant hosted Silmari for teams who do not self-host.
- Enterprise features. SSO, audit, compliance, workgroups, SLAs.
- FDO marketplace (optionality, later stage). Certified-FDO directory with take-rate.
§8.2 · Use of Funds
- ~50% engineering — multi-tenant plus auth, managed-cloud alpha, FDO training substrate
- ~20% first 5 senior FDOs (Landon-sourced from displaced-operator pool)
- ~10% ecosystem and open-source stewardship (MCP community, agent-builder partnerships)
- ~10% infrastructure and operations (hosted managed-cloud alpha, compliance readiness)
- ~10% founder runway and buffer (24 months full-time plus reserve)
§8.3 · Series-A Exit Criteria (24-month window)
- FDO time-to-competence under 10 weeks across 5+ hires (dogfood proof)
- 500+ paying seats at $150/mo blended = $1M ARR individual
- 20+ paying enterprise engagements on the coupling protocol
- Managed cloud in closed alpha
§9 · Glossary (Canonical Definitions)
- Memory Substrate
- Foundational, multi-tenant, infrastructure-level layer beneath any AI agent or AI tool that stores, organizes, and surfaces a user's accumulated context. Distinct from "memory layer" in that a substrate is presumed to underlie every consumer of memory rather than being one option among peers.
- Coupling / Decoupling Protocol
- Silmari-specific MCP-native mechanism that allows an individual operator's personal memory graph to be temporarily coupled to an enterprise federation graph during an engagement, then cleanly decoupled (revocable, auditable, bounded) when the engagement ends.
- Forward Deployed Operator (FDO)
- A senior operator hired from the displaced-knowledge-worker pool, trained on the Silmari method via the 12-week premium-onboarding curriculum, then deployed into Shape 2 and Shape 3 client engagements. Carries a portable Silmari substrate across engagements.
- Domain Encoding (Layer 01)
- The implicit knowledge a user teaches their AI tools across hundreds of conversations without realizing it: jargon, abbreviations, default assumptions, domain-specific reasoning patterns.
- Workflow Calibration (Layer 02)
- The user's procedural defaults, decision patterns, style, and operating cadence. Encoded through repeated interaction.
- Artifact / Capability (Layer 03)
- The provenance trail of what the operator made, how it was made, and why it was judged good. Notably absent from every competing AI memory platform.
- Harness Company
- Aubakirova and Bornstein's April 2026 a16z framing for memory-layer startups (Letta, mem0, Subconscious) that build orchestration and scaffolding around the context window. Used pejoratively. Silmari positions explicitly outside this category by claiming the substrate beneath rather than the harness around.
- Filing Cabinet Fallacy
- Same Aubakirova / Bornstein essay's attack on retrieval-based memory: "a bigger filing cabinet is still a filing cabinet; retrieval is not learning." Silmari's seven-question moat answers this directly via two rows (gets denser with use; structure emerges from use) that no filing cabinet can satisfy.
- Coupling Fee
- Silmari's enterprise revenue mechanism: per-seat or per-engagement access charges levied when an enterprise federation graph couples to an individual operator's substrate.
- Chain-Compounded Accuracy Baseline
- The realistic per-step accuracy of a 20-step LLM pipeline at the industry-standard 95% step accuracy: 0.95 to the 20th power, approximately 36%. Silmari production workloads run at 87% on the same workload class, a 2.4x improvement.
- MCP (Model Context Protocol)
- Open protocol introduced by Anthropic enabling structured tool, resource, and context exchange between AI clients and external services. Silmari is MCP-native: every Silmari capability is exposed as an MCP tool, allowing any MCP-compliant AI client to read from and write to a user's substrate.