Skip to main content

daimon

Your agent forgets everything between sessions. daimon writes a small checkpoint when a session ends and turns it into a briefing when the next one starts — so the agent resumes from a faithful prior state instead of a confident guess:

While you were away — here's where we left off.

Decisions made:
- [✓ verbatim] Postgres advisory locks over Redis locks for the scheduler  — "let's not add a Redis dependency for this"

VERIFY BEFORE TRUSTING (state may have changed outside this session):
- [✓ verbatim] PR #212 state — you said you'd merge it yourself from the UI  — "I'll merge it after the demo"

Open loops:
- [✓ verbatim] Retry policy for the payments webhook — exponential or fixed?  — "don't ship the retry loop until we pick a policy"
- [~ inferred] The staging config drift needs an owner [carried]

Active topic: Migrating the scheduler off cron to the new worker pool

Three nouns cover the whole system. A session is one conversation with your agent. A checkpoint is the local, signed record a session leaves behind when it ends. A briefing is the skimmable rendering of the latest checkpoint, injected when the next session starts.

Why it's different​

Most memory tools store what a model wrote about what happened — and when that text is wrong, nothing tells you. daimon marks every line with how it came to exist: [✓ verbatim] lines are exact quotes a deterministic verifier found in the session transcript (pure string operations, no LLM — case, whitespace, and formatting folded, elided spans allowed); [~ inferred] lines are the model's own conclusions, honest about being derivations. A "quote" that fails the check is demoted to inferred on the spot, so a hallucinated quote can never wear the verbatim badge. With receipts enabled, the checkpoint's exact bytes are signed — everything above is checkable offline, without trusting daimon, the model, or this page.

An independent reading​

Agent Memory Atlas, a commit-pinned survey of agent memory systems, has read this codebase several times since July 2026. From its latest executive summary:

This is one of the very few that treats the model's own trust label as a claim to be falsified before it is stored.

It is not praise alone. The same summary names the weak half: "The retrieval half is much less proven." Both halves are checkable against the card's file and line citations, and every reading is recorded with the commit it was pinned to.

The loop​

  1. A session ends — a host hook serializes the transcript into a checkpoint. Local JSON on your disk; no server, no telemetry.
  2. The next session starts — a hook injects the briefing as context, so the agent answers "where did we leave off?" before you ask.
  3. You verify, resolve, or correct — open items carry forward until closed, and go visibly stale instead of lying forever.

Start with the Quickstart — install to first briefing in five steps. Per-host setup lives in Hosts; the trust system is explained in Trust classes.