Anamnesis — Product Datasheet
Anamnesis is a persistent memory layer for AI tooling. It captures what your team and its AI assistants learn while working, consolidates it into trustworthy long-term memory, and serves the right context back — to any model, in any session, current as of the last decision.
What it does
| Capability | What you get |
|---|---|
| Automatic capture | Sessions are captured as they happen — no note-taking ritual, no "please summarize" step. Content is screened at the door; junk and injection attempts don't become memory. |
| Consolidation | Raw capture is distilled and organized into durable memory, not an ever-growing transcript pile. The result stays small enough to inject and cheap enough to search. |
| Cross-model recall | One memory pool serves every connected client. A decision made in one tool is known in the next — across model vendors (measured). |
| Provenance & trust | Every memory knows where it came from. Unvetted content is labeled and kept out of decision-weight, so imported or generated text can't impersonate your team's decisions (measured). |
| Belief revision | When a decision changes, the old answer retires from current context automatically — and stays available for audit and "what did we believe then" queries (measured). |
| Dashboard | Browse, search (semantic + date-aware), review flagged contradictions, and manage memory from the web — without the server reading your content to render the view. |
What it replaces
- Re-explaining project context at the start of every session — the search tax: measured at ~5× more turns and ~5× slower answers without memory (T5).
- Wiki pages and CLAUDE.md files that go stale silently.
- Naive RAG memory that degrades as it grows and serves superseded decisions as current (T2, T3).
Integrations
- Claude Code — first-class plugin: capture and recall ride the session automatically.
- MCP clients — Claude Desktop, claude.ai, and any Model Context Protocol client connect to the same pool.
- Multi-model — in daily use across Claude Code, Google Antigravity, and Codex against one shared pool.
Deployment
| Tier | Description | Status |
|---|---|---|
| Hosted | Managed service, per-user encryption at rest, TLS in transit | Live (private beta) |
| Client-held keys | Server stores ciphertext it cannot read | Published roadmap |
| On-premises | Fully local: memory, consolidation, and inference on your hardware, with one-way-mirror read access to shared pools | Design partner conversations |
Performance
Published, reproducible, blind-graded benchmarks — not marketing numbers: near-oracle answer quality at an ~18× token discount, fidelity maintained at 1,000-memory scale where naive retrieval collapses, verified cross-model handoff. Full method and results: benchmarks.
Security
Per-user encryption with no master key, scoped read-only tokens, screened capture, published threat model. Summary: Trust Center · full detail: security page.
Getting started
Private beta. hello@jiashley.com — tell us about your team and stack; design-partner conversations welcome for the on-premises tier.