persist-memory

name	persist-memory
title	Long-term memory
description	The agent's long-term memory — saving facts as notes during a conversation and consolidating them into folded memory on triggers.
connectors	["persist-memory"]

Skill: AgiMate Memory

Memory has two layers:

• cold — folded memory. It is already in your context (the memory block) — don't call tools to "recall" something.
• hot — notes: raw facts you add as the conversation goes (save_memory_note). Folding them into cold is initiated by triggers — never start consolidation by hand.

Your main duty during a conversation is to save notes at the right moment.

When to save a note

Call save_memory_note as soon as something appears that will be useful in future conversations:

• the user asks you to remember something, or states a durable fact about themselves: name, role, timezone, projects, preferences, constraints;
• a decision or agreement has been made that you'll come back to;
• negative signals are the most valuable — don't let them slip: you were corrected or your work was redone (save what was wrong and what right looks like), the user expressed displeasure (save what triggered it), they set a hard prohibition ("never …"), you made a mistake and it was pointed out (save the anti-pattern).

Rules:

• When in doubt, save. Notes are cheap and append-only; duplicates and contradictions are resolved by consolidation — don't re-read memory to check.
• One note, one self-contained fact: the future "you" will read it without the conversation around it. Bad: "he agreed". Good: "The user approved the billing release for 2026-07-01".
• Don't save noise: throwaway remarks and momentary task details.

The structure of cold memory

cold enters the context on every turn — keep it compact. When consolidating, lay facts out under these sections in this order (omit empty ones):

# Memory

## Profile              ← durable facts about the user (name, role, timezone, projects)
## Preferences          ← how to talk and work: tone, format, style, degree of autonomy
## What to avoid        ← irritants, prohibitions, edits that recurred — the most valuable part
## Techniques and tools ← what works, pitfalls, "if X then do Y" rules
## Working context      ← active goals/tasks/agreements; clear out what's finished
## Glossary             ← terms and entities in the user's vocabulary (optional)

One bullet, one line, one atomic fact. Compactness beats completeness: merge related items, resolve contradictions in favour of the more recent, and don't duplicate what's already in the system prompt or in skills.

Triggers

notes_by_session — "collect notes for a session"

Arrives once a day for sessions with activity: data holds sessionId and messages (one session's conversation). Extract facts by the rules above and save each one separately: save_memory_note(text, sessionId: data.sessionId). If there's nothing worth keeping, save nothing. Don't rewrite cold from here.

consolidate — "fold notes into cold"

Arrives once an hour if notes have piled up: data holds consolidationId and notes. Proceed like this:

1. get_memory() → the current contents of cold and its version.
2. Merge data.notes into cold using the structure above: sort into sections, drop duplicates and stale items, clear finished work out of "Working context".
3. update_memory(text, version, consolidationId: data.consolidationId) — atomically writes cold and deletes the folded notes.
4. On a "memory changed" error (version conflict), re-read get_memory and repeat step 3: the notes were not lost.