Conversation memory
Skill Amey-Thakur/AI-SKILLS/skills/ai-memory-rag/conversation-memory
Keep a long conversation coherent within a finite window through rolling summaries, pinned facts, and selective recall. Use when sessions run long enough that early context falls out.From its SKILL.md
npx -y skills add Amey-Thakur/AI-SKILLS --skill conversation-memoryAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 5 stars5 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
1.8 KB, as published. Nobody here has run it
Conversation memory
Every long conversation eventually exceeds the window, and the naive answer of dropping the oldest turns loses exactly what was established first: names, decisions, and constraints. Memory is choosing what survives.
Method
- Separate durable facts from conversational flow. Decisions, preferences, and constraints are pinned; small talk and superseded attempts are not.
- Summarise rolling history rather than truncating. A running summary of older turns preserves the thread at a fraction of the tokens, and truncation loses it entirely.
- Keep the recent turns verbatim. Recency matters for coherence, so the last several exchanges stay unsummarised while older ones compress.
- Re-summarise incrementally. Summarising the summary each round compounds distortion, so summarise from the source turns where possible.
- Extract facts explicitly as they are established. A structured note when the user states a constraint is more reliable than hoping a summary preserves it (see long-term-user-memory).
- Handle contradiction by recency with acknowledgement. When the user changes their mind, the new statement wins and the change should be visible rather than silent.
- Show what is remembered. Users are unsettled by a system that remembers invisibly, and a visible memory is also a correctable one.
Boundaries
Summarisation loses detail by design, and the lost detail is sometimes what mattered. Compression costs a model call per cycle. Conversation content is user data with retention and deletion obligations (see right-to-erasure).