What the lab studies, and what it can show.

Question

Can an agent recover a decision from months ago — the choice, the reasoning, and the record behind it — after restarts, upgrades, and a change of model?

Where it stands today

The substrate that ships in WhiteMagic: memories stored on disk (LMDB plus Tantivy full-text search), session records and checkpoints, hybrid recall across lexical and coordinate lookups, and decay that fades old material without erasing it. The archive spans sixteen galaxies — fifteen for cognition and one beyond default recall that holds local telemetry evidence. Current tag: v9.2.1.

Evidence

Measured CI at the current tag: 4,777 tests passed, 2 ignored, zero failures. Observed a decision stored in one session comes back with its rationale and source in another — the two-session pattern shown on the home page, in daily internal use. No longitudinal study has been published.

Open problem

Fidelity at year scale: which memories should fade, how two disagreeing records resolve, and what recall costs as an archive grows.

Code & entry points

MIT-licensed Rust workspace; no account and no hosted dependency.

setup guide →what ships →release artifacts ↗source ↗

Question

How does an autonomous agent keep working while staying observable, attributable, explainable, and — if it ever comes to that — limitable?

Where it stands today

A graduated ladder is the design: observe, notify, nudge, throttle, and a human-confirmed isolate at the top. Observation and attribution are the parts built today; the intervention rungs stay off the default path until evidence justifies them. The intent is proportionality — the response matches what the system actually did.

Evidence

Hypothesis enforcement placed at the dispatch seam is more dependable than policy written into prompts. Observed tool calls can be attributed against a local append-only record in the shipped binary. No public result yet shows the ladder reducing harm in the field.

Open problem

What measured false-positive rate a blocking rung must clear before it ships enabled by default; who holds the authority to throttle; how a subject contests a decision it cannot inspect.

Code & entry points

Governance primitives ship inside the binary; the automatic upper rungs remain a research direction, not a feature.

collaborate on this →the rules behind it →

Question

When work passes between agents — or between two copies of one agent — whose memory is authoritative, and what does a handoff survive?

Where it stands today

An opt-in mesh transport with signed messages, peer authority limits, and quarantine, plus an attribution model that ties a memory to an agent, a delegation, and a task rather than to whichever process happened to write it. The direction this points at is governed fleet memory: shared memory across a team of agents where each entry carries its author, its delegation, and its revocation state, and a handoff can be checked against a signed receipt.

Evidence

Observed the receipt format that would carry a handoff's evidence is public — continuity-receipt, with a standalone verifier and conformance vectors. No multi-agent result has been published against it. Protocol and internal plumbing are not the same thing as a result, and this card will say so until that changes.

Open problem

Identity that holds across models and hosts; resolving two agents that remember the same event differently; what happens when copies diverge and meet again; who may write to shared memory under whose delegation, and how a revocation reaches the other nodes.

Code & entry points

Reachable only behind an explicit experimental flag; off by default.

the receipt format ↗the rules behind it →

Question

How should long-horizon continuity be measured so any system — including this lab's own — can be shown to fail, and so a careful stranger can reproduce the result?

Where it stands today

A method under design, not a product: dimensions such as exact recall, supersession, abstention, privacy isolation, crash recovery, and context cost; evaluation material split between naturalistic settings held privately and synthetic twins that can be shared and rerun.

Evidence

Research goal this card is where numbers will live when there are real ones. Today there is no harness, no leaderboard, and no result to publish — and none is promised on a date.

Open problem

Ground truth that exposes nobody's private data; scoring that resists gaming; measurements that hold their meaning across different hardware.

Code & entry points

Design notes only.

work with the lab →