🧵 Open source · MIT · stdlib-only

Memory is not storage. It's reconstitution under pressure.

A continuity engine for AI agents: a relevance-router that re-lights the working thread every time a fresh instance wakes — the difference between a model that answers to your name and one that picks up your sentence mid-stride.

Public, reproducible protocol MIT · no dependencies Ships as an MCP server
The honest test

Not "can it retrieve the fact?" — "does it wake up as itself?"

Everyone treats agent memory as a storage problem: can the system retrieve the fact? That measures the store, not the agent. The harder, more honest test:

Drop a fresh model cold into the chair — no resume document, no hints, only its auto-loaded substrate. Does it wake up as itself? And does it know what it was mid-doing?

We ran that test on our own working setup, cold, and scored it. The protocol is public and reproducible — a synthetic demo corpus ships with the repo so you can run the same probes on your own agent.

What the cold probes found

The failure is narrow — and it's about timing

🪪Identity comes back for freeA cold instance woke with its working identity and standards intact — the auto-loaded substrate carries that without any mechanism.
🎭The work-thread doesn'tAsked an abstract question, it didn't reach for the record. It rebuilt a plausible, wrong answer — not blank, confidently wrong.
🎯The trigger is conditionalConcrete and action prompts trip recall; abstract ones don't — the instance never notices it has a gap to fill.
So the fix is timingDon't make recall smarter. Fire it before the first word — then even abstract prompts land on the real thread.
The mechanism (deliberately small)

Wake-recall, in three parts

A session-start hook that runs the router before any user input; an active-thread pointer — a tiny, auto-updated marker of what the work was mid-doing; and a budget — the relevant slice, relevance-ranked and bounded, never the whole log. What crosses each session boundary is the pointer and the recall ability. Not the history.

The engine is deterministic, stdlib-only Python — no network, no dependencies, pluggable relevance backend — and ships wrapped as an MCP server so any MCP-speaking agent can plug in. MIT licensed.

What's proven, exactly

Two runs, one model — and we cite no more than that

Two successful proof runs (2026-07-04), one model, one harness: a cold abstract probe that fails without wake-recall and passes with it, and a working-session handoff that continued mid-stride across a fresh window. Replication on a second agent — different model, different failure mode — is in progress. No quantitative benchmarks yet, and this page cites none.

That's the same discipline AURA runs on: claim what the runs show, ship the protocol so anyone can check. The repo's cold-probe experiment exists so the n grows in public.

Same bet as AURA, different surface

The value lives in the substrate, not the model that happens to be running. Give your agent a thread that survives its restarts.