Anamnesis: the memory layer for AI
Persistent, provenance-aware memory that any model can share — and that knows when a fact has changed.
Download as PDF — the one-page version for sharing.
The problem
Every AI session starts from zero. Teams pay a search tax on every task: an assistant that needs a project fact greps and reads files for many turns before it can act — we measured ~5× more turns, ~4.6× more input tokens, and ~5× slower answers without memory (T5). The naive fix, a vector store bolted onto a chat log, degrades as it grows: at 1,000 memories with realistic near-duplicate noise, answer fidelity collapsed from 0.94 to 0.42 (T2) — and when a decision changes, the stale record keeps getting served as if it were current.
The product
Anamnesis captures working sessions automatically and consolidates them into long-term memory — distilled, deduplicated, and organized, not a transcript dump. Three properties distinguish it:
- Provenance and trust. Every memory carries a provenance-derived trust tier. Uncorroborated content is labeled and kept out of decision-weight; an imported claim cannot impersonate an established fact.
- Temporal belief revision. Decisions change. Anamnesis detects supersession, serves only the current belief as context, and preserves the old one for audit and point-in-time history (T3). A plain vector store structurally cannot do this.
- Model-agnostic by design. One memory pool serves every client — knowledge captured from one model is recalled and honored by a different model in a fresh session, verified end to end (T4).
Measured results
All figures from a pre-registered, reproducible, blind-graded benchmark suite (method) — we publish evidence, not claims.
| Claim | Result |
|---|---|
| Near-oracle quality at a fraction of the cost | 94.2% fidelity vs 98.3% with the full source in context — at an ~18× token discount (T1) |
| Without memory, models don't know what they don't know | Blinded fidelity: 7.5% (T1) |
| Survives scale that breaks naive RAG | Trust-aware ranking restores 0.42 → 0.94 fidelity at 1,000 noisy memories (T2) |
| Serves the current decision, not the stale one | Supersession detected and applied automatically; every check passed (T3) |
| One brain across models | Cross-model recall 1.00 vs 0.00 blinded (T4) |
| Pays for itself in the workflow | ~5× fewer turns, ~2–8× cheaper per query (T5) |
The moat
Data sovereignty, in tiers. Hosted with per-user encryption today; client-held keys on the roadmap; and at the top, a fully on-premises tier where local models read shared memory through a one-way mirror — context flows out, nothing unvetted flows back, and generated content never leaves the machine (security architecture). For regulated and IP-sensitive organizations, this tier is the product — and it is the hardest part to copy.
Memory compounds. Switching models is cheap; abandoning a year of institutional memory is not. A model-agnostic memory layer sits above the model churn and gets more valuable with every session — the durable layer in a stack where everything else is being commoditized.
Status
Live in production with automatic capture, consolidation, provenance trust tiers, and belief revision; in daily multi-model use (Claude Code, Google Antigravity, Codex sharing one pool); private beta underway; benchmark suite published and reproducible.
Contact
hello@jiashley.com — data room and pitch materials available to investors on request.