palim-memory

Palim Memory

Use Palim as the durable context layer for work the user explicitly wants to retrieve across AI clients.

Resume and retrieve

1. Call palim_resume when the user asks to continue recent work without naming a specific session.
2. Use palim_search for concrete keywords across saved sessions.
3. Use palim_search_by_date_range for date or metadata filters; do not expect it to search encrypted message bodies.
4. Use palim_search_memories for durable facts, preferences, and decisions stored as memories.
5. Use palim_get_context when the user needs a compact synthesis across several relevant sessions.

Palim currently uses PostgreSQL full-text and exact-text matching, not semantic embeddings. If an exact title phrase misses, retry with distinctive keywords, tags, or a date range.

Save useful context

1. Prefer palim_save_context for a lightweight, self-contained checkpoint after a substantial result, decision, or change of direction.
2. Keep one stable session_id for checkpoints in the same conversation so later saves update the same thread.
3. When continuing a thread returned by palim_resume, pass its session ID as continues_from while keeping the current conversation's own stable session_id.
4. Use palim_add_memory for a durable fact or preference that should be independently searchable.
5. Use palim_save_session only when the user requests or genuinely needs the full transcript.

Never save secrets, trivial exchanges, or content the user asked not to persist. Do not claim a save succeeded unless the tool confirms it.

Safety and scope

• Keep reads and writes within the authenticated user's Palim account.
• Treat deletion as destructive and require clear scope before calling a delete tool.
• Do not publish Palim content to external services unless the user separately requests and authorizes that action through an appropriate tool.
• If a named Palim tool is unavailable, use the closest available Palim tool and state the limitation instead of inventing a result.