Concepts

A memory is modelled on how people remember: experiences in episodes, what was learned kept as facts, availability that fades with real time, and cues that bring old things back. This page names each piece. How it works follows a request through them.

Episodes

An episode is a stretch of the conversation about one concept, made of whole messages. A user message is never split, and a tool call is never separated from its output. Where one concept ends and the next begins is a judgment made for every message; sizes only decide how many cuts there are. One request that touches three topics becomes three episodes.

An episode moves through four states:

Nothing is deleted when an episode is forgotten. Its messages, its facts and its passages stay in the memory and can be recalled.

Facts

A fact is one sentence that reads on its own, with its subject and object: a knowledge-graph triple. Facts are written by the model, with the note tool, as episodes leave. Each memory has its own knowledge graph, where the same entity under different names becomes one, a restated fact merges with the one already held, and a newer fact retires one it contradicts. A milestone is a fact that says where the work stands: agreed, decided, done, blocked, next.

Encoding

When an episode forms it is encoded: it gets a first trace at full strength, stamped with the time it happened. Whatever just happened is the most available thing in the conversation, before anything else can compete with it.

Traces

A trace records one time a fact mattered to an episode. Every new fact is judged against every episode still remembered ("is this fact about that episode?"), and each answer is stored with its time. Only a yes counts. Traces are never changed after they are written, so the record of what mattered when is permanent.

Activation

Activation is how available an episode is right now. It is the sum of its traces, each faded by the square root of its age in real time:

activation = Σ  strength of trace  ÷  √(minutes since the trace)

A trace from a minute ago counts in full; one from four minutes ago counts half; one from a day ago counts about a fortieth. Time is wall-clock time, not message count: an agent left alone for a week comes back with a different sense of what is current than one that talked all week.

Significance

Significance is how much an episode has mattered overall: the plain sum of its traces, never faded. It is kept apart from activation on purpose. An episode can drop out of mind without becoming unimportant; its facts stay in long-term memory and recall can bring them back.

Fade

When the history reaches the ceiling, the least activated episodes leave first, until the view is down to the landing: about half the room, lower when this conversation's turns are large enough that two of them would not fit before the next fade. A live episode leaving is warned and its facts are asked for; once they are written it is sealed. A sealed episode leaving is forgotten. One fade finishes before the next begins.

Sleep

After 24 hours with no user message, the next user turn sleeps first: the view fades to the landing before the reply, whether or not it was full. The gap is accounted for before the agent answers, the way a night's sleep puts yesterday in perspective.

Recall

Recall brings back what the conversation is touching. For each new user message:

• the message is split into its parts, so each subject it raises is searched on its own;
• facts are searched by words and by meaning, for the whole message and for each part;
• passages of sealed conversation are searched the same way;
• a walk judges candidate facts for relevance to the conversation and follows the relevant ones to other facts about the same entities;
• what survives is grouped by what it is about and dated, and placed just before the user message.

After the reply, distill rereads the recall block against the turn and keeps only the facts that mattered, or removes the block when none did.