mnemon

mnemon — Persistent Memory

Memory Stores

• Private (default): Per-group, read-write. All writes go here.
• Global: Shared across all groups, read-only. Use --store global --readonly to query.

Recall — before responding

Default: recall on every new user message, unless ALL of these apply:

• Direct follow-up within a topic already fully in context
• No reference to past sessions, decisions, or preferences
• No knowledge dependency beyond the current conversation

To recall:

mnemon recall "<query>" --limit 5
# Also check shared knowledge:
mnemon recall "<query>" --store global --readonly --limit 5

Craft a focused, keyword-rich query — do not pass the raw user prompt.

Remember — after responding

Run this decision tree after every substantive response:

Step 1 — Does this exchange contain any of these? a) User directive — preference, decision, correction, explicit "remember this" b) Reasoning conclusion — non-trivial judgment from multi-source synthesis c) Durable observed state — system fact, environment detail, architectural finding → No to all → STOP.

Step 2 — Does a highly overlapping memory already exist? → Yes, incremental new info → STORE the new content; review the existing memory separately → Yes, but contradicts/supersedes → STORE and verify the new fact; explicitly forget the old ID only if superseded → No significant overlap → CREATE

Step 3 — Is it worth storing? Rebuilding from scratch costs more than storing + recalling?

• Single-query public facts → No
• Multi-source synthesis with non-obvious conclusions → Yes
• User-specific context no search engine can recover → Yes → No → STOP.

What to store: conclusions and user-specific context, not raw facts.

Workflow

1. Remember: mnemon remember "<fact>" --cat <cat> --imp <1-5> --entities "e1,e2" --source agent
   • Only exact content repeats are skipped; distinct content is stored and diff suggestions are advisory.
   • To retire a superseded memory, store and verify the new fact, then explicitly run mnemon forget <old-id>.
   • Output includes action (added/skipped), semantic_candidates, causal_candidates.
2. Link (evaluate candidates from step 1 — use judgment, not mechanical rules):
   • Review causal_candidates: does a genuine cause-effect relationship exist? causal_signal is regex-based and prone to false positives — only link if the memories are truly causally related.
   • Review semantic_candidates: are these memories meaningfully related? High similarity alone is not sufficient — skip candidates that share keywords but discuss unrelated topics.
   • Syntax: mnemon link <id> <candidate> --type <causal|semantic> --weight <0-1> [--meta '<json>']
3. Recall: mnemon recall "<query>" --limit 10