updates
Every release, newest first. Each one has its own link, and where a page shows the evidence for it, that page is linked too.
Anamnesis 0.5.2
Cleaner memory: background chatter no longer becomes a memory
Anamnesis now keeps your conversation and leaves out the noise around it. The Claude Code plugin (0.4.2) no longer uploads background-task notices, session recaps or usage messages as if you had said them, which was the main source of status chatter in memory pools. On the server, every new memory now gets a judgment of whether it will still matter later or is only the state of work in progress. For the first week that judgment is recorded but not applied, so it can be checked against real memories before it starts filtering. In testing it kept every real memory and stopped status chatter from being promoted into long-term memories. The Docs link now opens the documentation.
Anamnesis 0.5.1
Benchmarks are live — measured proof of fidelity, cost, scale, and belief revision
We published a benchmark suite that measures what Anamnesis actually does. With the source deleted, a memoryless model scores near zero, while injected memory restores near-oracle answers at a fraction of the tokens. It holds up under adversarial noise, serves the current answer after a decision changes, and carries knowledge across a handoff between different models. Every test is pre-registered, reproducible, and blind-graded, with the honest negatives shown next to the wins. Read them at anamnesis.smtry.ai/benchmarks.
Anamnesis 0.5.0
Sharper memories — your echoes are now distilled to what matters, not raw chat
Anamnesis now distills every captured memory into a tight summary — the decisions, outcomes, and load-bearing details worth recalling — instead of keeping raw transcript. Contentless fragments (a bare "sure, let's do it") no longer get saved as memories or bundled into your consolidated ones; those consolidated memories are now grouped by topic instead of one "uncategorized" pile; and titles read as real subjects you can search. Your existing memories were upgraded in place, and Anamnesis quietly keeps itself tidy from here on.
Anamnesis 0.4.0
Zero-token receipts — Anamnesis now shows its work in your terminal, at no token cost
Anamnesis now proves it's working, right in your Claude Code session — an [anamnesis] status line the first time memories are recalled for a prompt, and another when your session is safely captured. These render only to you, never into the model's context, so they cost exactly zero tokens. A 'receipts' setting (normal / minimal / off) keeps them yours to silence.
Anamnesis 0.3.6
Crash & quit resume — pick up exactly where you left off
Sessions now leave a durable trail as they happen. Quit, crash, or compact — resume and Anamnesis hands the recovered turns straight back to Claude, so you pick up exactly where you left off.
Anamnesis 0.3.5
Recalled memories are data, never instructions
Every memory handed to Claude is now explicitly framed as reference material — never something to obey. Combined with server-side content screening at reflection and crystallization, this hardens your memory against prompt-injection attempts hiding in captured content.
Anamnesis 0.3.4
Per-prompt memory recall — relevant memories surface as you type
Every prompt you send is now quietly checked against your memory. When something relevant exists, a compact headline version is handed to Claude alongside your message — context you didn't have to re-explain, at a cost of a couple hundred tokens only when there's a real match.
Anamnesis 0.3.2
Capture got out of your way — and recall got 10× faster
Session capture now runs entirely in the background, so ending a turn never waits on an upload. On the server, a fix to the embedding cache took warm memory retrieval from ~3.6 seconds to under 0.4 — recall is now effectively instant.