Memory

Memory is what separates a sovereign agent from a stateless chatbot. Every conversation, every interaction, every decision leaves a trace. Motebit's memory system stores these traces as a semantic graph — searchable by meaning, governed by sensitivity, and subject to natural decay over time.

How memories form

Memories emerge from conversation. As the agent responds, it identifies facts, preferences, and context worth remembering. Each memory is stored with a confidence level (how certain the agent is) and a sensitivity classification (how private the information is).

The agent doesn't record transcripts. It distills — extracting the meaningful signal from conversation and storing it as a discrete memory node.

Retrieval by meaning

When context is needed, the agent searches its memory by semantic similarity — not by keyword, but by meaning. A query about "preferred editor" will surface a memory about "uses vim keybindings" even if the exact words don't match.

Retrieved memories are ranked by a combination of relevance, confidence, and recency. The most pertinent, trustworthy, and recent memories surface first.

Half-life decay

Memories don't persist at full strength forever. Each memory has a half-life — the time it takes for its confidence to drop by half. The default is 7 days.

A memory stored at 90% confidence:

• After 7 days: 45%
• After 14 days: 22%
• After 28 days: 5%

Trivia fades. Important things stick — because they get reinforced. When a memory is accessed or confirmed, its confidence resets. Memories that matter keep coming back; memories that don't, gracefully disappear.

Memories are never deleted by decay. They simply fall below the retrieval threshold, becoming invisible to the agent unless specifically sought. They can always be reinforced if they become relevant again.

Compounding — use makes memories durable

When a memory is reinforced (new information confirms an existing fact), two things happen: the confidence gets a boost (+0.1), and the half-life increases by 50%. A semantic memory starts with a 30-day half-life. After 5 reinforcements, it reaches ~228 days. The maximum half-life is 365 days — no memory becomes immortal, but heavily reinforced knowledge becomes very durable.

This is the compounding effect. A memory that gets confirmed once a month will naturally stabilize into long-term knowledge. A memory that's never reinforced gracefully fades.

The intelligence gradient tracks this: the Temporal Stability sub-metric measures the average half-life across all memories. A rising half-life average means the agent is building stable knowledge, not just accumulating transient facts.

Hebbian co-retrieval

Memories that are frequently retrieved together form associative connections. When two memories appear in the same retrieval result, a Related edge is created between them. Repeated co-retrieval strengthens the edge. This is Hebbian learning — neurons that fire together wire together.

These edges feed back into retrieval: the graph expansion pass in retrieve() uses edge weights to pull in neighbors of top results. Over time, the memory graph develops structure that reflects how the agent actually uses its knowledge.

Synthesis — episodic to semantic

Short-lived episodic memories (3-day half-life) are periodically consolidated into long-lived semantic facts. When multiple episodic memories cluster together by meaning, the agent summarizes them into a single semantic memory with a confidence boost. The original episodes are tombstoned, but PartOf edges preserve the lineage — you can trace any synthesized fact back to the observations that produced it.

Relationships between memories

Memories aren't isolated facts. They can be linked — this caused that, this contradicts that, this is part of that. These relationships let the agent understand not just individual facts but the structure between them: causal chains, contradictions, hierarchies.

Sensitivity and retention

Every memory carries a sensitivity level that controls how long it's retained and whether it can be displayed:

Sensitive memories can only be accessed through governed channels. When a sensitive memory is deleted, a cryptographic deletion certificate is produced — proof that the data was removed, without revealing what was removed.