Observation-first memory kernel · Python 3.11+

Preserve the evidence.
Derive the belief.

An observation-first kernel that preserves evidence and derives current beliefswithout overwriting history. Immutable observations, a confidence resolver, and a repository graph you can interrogate.

GRAPH 2,680 nodes 6,029 edges 168 communities AMBER = EVIDENCE · TEAL = INFERENCE
01 — THE MODEL

History is never overwritten

Most memory systems mutate: new information replaces old, and the trail is gone. Anamnesis keeps the two layers apart — the observation is immutable and permanent; the belief is derived, current, and always revisable as evidence accumulates.

● Observation — evidence

What was seen

Immutable, timestamped, never edited. The record of what actually happened stays intact no matter what is later believed about it.

● Belief — inference

What is currently held

Derived by a confidence resolver from all evidence to date. Revised, not overwritten — and always traceable back to the observations that support it.

02 — THE SQUAD

Four roles, one memory

Validated desktop agent profiles, each with a distinct lens on the same evidence. They request no microphone, speaker, display, network, or model access unless a profile grants it.

Profile

Scout

Notices what changed — surfaces new observations against the standing record.

Profile

Witness

Compares claims against evidence, holding beliefs accountable to their support.

Profile

Dreamer

Proposes testable hypotheses — inference reaching past the current evidence.

Profile

Librarian

Governs archive metadata through a domain-agnostic boundary; retrieves on request.

03 — THE RESEARCH

A result that reversed itself

A public, model-free measurement track (built on LongMemEval) asked a sharp question about delegation — whether one agent handing retrieval to another's lens actually helps. The honest answer wasn't the expected one.

Hypothesis
Delegating retrieval to the Librarian's lens should improve answers across the board.
What the numbers showed
The delegating run retrieved better on every measure — yet answered no better. The cause was prompt size, not retrieval: aggressive summarisation threw away the gain.
The finding
Delegated retrieval works; delegated summarisation can erase it — delegation pays exactly when the delegating agent's own lens is retrieving badly.

The point isn't a smarter answer. It's that every belief can be walked back to the evidence it stands on — and the evidence is still there.