Anamnesis Live · free alpha

One memory,
read by every AI tool you use.

Say a thing once and it is there in the other tools. Encrypted with a key derived from your own credential, with no master key. Free while it is in alpha.

Try the alphaHow it protects you →

In practice

Said on Tuesday in one tool.
Back on Thursday in another.

In the terminal tools, what you say while working is captured as you work. In the chat apps, the assistant saves what you ask it to keep. Either way it comes back where the next job is.

claude code · ~/SMTRY
$ claude
› pick up the onboarding doc where we left off
[anamnesis] recalled 3 memories for this prompt: context you didn't have to re-explain.
[decision · Tuesday · Claude Desktop]
The onboarding doc is for engineers on day one, not managers; keep it under two pages.
[preference · Tuesday · ChatGPT]
Second person, present tense, one command per step, no screenshots.
[reference · last week · Codex CLI]
The setup script lives in tools/bootstrap.sh and takes about four minutes.
Claude: Picking up the draft with those three in hand. Section 2 is the one still open.
What a recall in Claude Code looks like, drawing on things said in Claude Desktop, ChatGPT and Codex. Try it free, in alpha →

Measured

42 ms

Retrieval, 13 to 42 milliseconds across the scale benchmark, flat through 1,000 records. The benchmark →

18×

Fewer tokens than pasting the source. Injected memory reached 94.2% of the oracle answer quality at about one eighteenth of the tokens. How it's measured →

7

Tools, one connector. One account, one memory, read by every assistant you already open. The architecture →

How it works

One thing you said, becoming memory.

  1. Session

    One conversation with a connected tool. In Claude Code, Codex CLI and Gemini CLI it is captured turn by turn as it happens; in the chat apps the assistant relays what was said. No screening has happened yet. The raw material is what was said, with the context it was said in.

  2. Capture

    Each capture is stamped with its time, its tool, and who said it. “Keep it under two pages.” “Second person, present tense.” “The setup script takes four minutes.” In the terminal tools the turns arrive as written; in the chat apps they arrive as the assistant relays them. The falling matter is the conversation; the gas is its context.

  3. Episode

    One thing the session was about: a decision made, a problem solved, a preference stated. A session is cut into episodes, each with a beginning, an end, and a centre of its own.

  4. Reflection

    An episode is read with what is already known in view, and what it meant is written down. In the science's words, elaborative encoding: new information tied to the structures already there. Gravity pulls the matter into a disc; the disc ignites.

  5. Echo

    A memory. The meaning of an episode, distilled. Four things orbit it: what was said, where it was said, when, and what kind of thing it is: a decision, an insight, a reference, a preference. The small moons are the details it keeps hold of: a path, a quoted line, a name.

  6. Crystallization

    Echoes that keep agreeing, across sessions and across tools, consolidate into something durable: memory consolidation, in the science's words. When a newer echo contradicts an older engram, the newer one, once confirmed, takes its place; the older one is not erased, it cools and moves aside. The star settles into a cluster; the one it supersedes reddens and drifts to the edge.

  7. Engram

    Long-term memory: what has been confirmed enough to be trusted. Engrams gather by meaning into convergence zones, so a task lands in a neighbourhood of knowledge rather than a flat list. The cluster is one such neighbourhood.

  8. Time

    Once something is a memory, the system does not delete it. When it stops being confirmed it cools and goes quiet, and keeping it costs nothing. The star swells, reddens, and lets its layers go.

  9. Dormant

    Still there, whole, waiting on a cue. The archive can grow without the cost of recall growing with it, because what is dormant is never loaded.

  10. Ecphory

    Recall triggered by a cue. It exists because loading everything would cost more with every memory kept; instead a new conversation's context is the cue, and only the memories that bear on it light up and are placed in front of the model. In the scale benchmark that read takes 13 to 42 milliseconds; the model's own reply time is separate. The cue reaches the gas; the gas falls back in.

  11. Recalled

    Back, in a different tool on a different day. In the terminal tools it arrives with the prompt; in the chat apps the assistant asks for it at the start of a conversation. Memories that do not bear on the question are not loaded.

Architecture

One memory pipeline, grounded in memory science.

One memory layer, seven clients on top of it, your keys underneath. If it is drawn, it runs.

reflection elaborative encoding crystallization memory consolidation SESS. SESSION A conversation, captured as it happens EPIS. EPISODE episodic framework, Tulving A topic-bounded unit of experience ECHO ECHO episodic memory trace Contextually bound. Unconsolidated echoes go dormant, never deleted insight what was learned decision what was chosen, and why reference facts worth keeping at hand preference how you like things done ··· more echo types E ENGRAM long-term memory substrate gathered into convergence zones Retrieved by context, never front-to-back. Only what bears on the task at hand. ecphory recall by cue, Tulving The next conversation is the cue; only what bears on it is recalled
SessionA conversation with any connected AI tool, recorded as it happens. Nothing is summarized yet; this is the raw experience.
EpisodeThe session is cut into episodes, each about one thing. Tulving's episodic framework: what happened, when, in what context.
EchoReflection turns an episode into echoes: typed traces such as a decision, an insight, a reference, or a preference. Echoes that never crystallize go latent; a later cue can bring one back.
EngramEchoes that keep being confirmed crystallize into engrams, the long-term substrate. Semon's word, Tonegawa's lab work.
Convergence zoneEngrams mass into convergence zones by meaning, so a task lands in a neighborhood of knowledge rather than a flat list.
EcphoryTulving's term for recall triggered by a cue. When a new conversation begins, its context is the cue: the engrams and echoes with bearing on it are activated and injected, and the rest stay dormant. The archive can grow without the cost of recall growing with it.

Sources. Tulving, E. (1972), "Episodic and semantic memory," in Organization of Memory · Müller, G. E. and Pilzecker, A. (1900), Experimentelle Beiträge zur Lehre vom Gedächtnis · Semon, R. (1904), Die Mneme · Tonegawa lab, MIT, engram-cell studies. The measured figures on this site link to the product's benchmarks page. What is drawn is what the stages are; the thresholds and rules that decide what consolidates and what is trusted are not, by design.

Retrieval

Memory with a time, a source, and a cue.

Every memory keeps its time, its source, and a link to whatever superseded it. Retrieval starts from the prompt in front of you, which is the cue: the memories most relevant to it, ranked by relevance, recency and trust, come back, and superseded ones are held back. In the scale benchmark that read takes 13 to 42 milliseconds, flat through 1,000 records. Over time, memories are distilled; in the fidelity benchmark, injected memory used about one eighteenth of the tokens of pasting the source at 94.2% of the oracle answer quality.

The Anamnesis dashboard
The dashboard's counts and live event stream. Memory bodies are decrypted only when you open one.

Yours

Every memory is yours to see, search and delete.

Each account's memory is encrypted with a key derived from its own credential; there is no master key. Export is plain JSON, anytime. It is in alpha: it runs and it is still changing.

Try the alphaHow it protects you →

Security and evidence

Where things stand.

Each account's memory is encrypted with a key derived from its own credential. Every memory is visible to you, searchable, and deletable. Captured content is screened before it becomes memory. We are not SOC 2 audited. Weekly operating evidence has been collected and committed since 20 August 2026; the audit starts when it is funded.

Read the security page →

The Anamnesis dashboard: tokens saved, tokens paid, retrievals, crystallizations and memories, with the current release note
The dashboard's counts. The memory list, searchable and deletable, sits below it.

Start with the free alpha.

One account, seven tools, your own keys. No card.