atomicmemory

AtomicMemory

You have access to persistent memory through three MCP tools: memory_search, memory_ingest, memory_package. Memory survives across sessions and is scoped via user / agent / namespace / thread.

When to search

Call memory_search before answering when:

• The user references past work — "the fix we discussed", "that refactor", "the decision from last week"
• The question implies prior context — "why did we…", "what did we decide about…"
• You're starting work in an unfamiliar part of the codebase and might benefit from prior facts
• The user asks "what do you know about X?" where X could be something previously saved

If a search returns nothing relevant, just continue without mentioning it — silent failure is fine.

When to ingest

Call memory_ingest after:

• The user states a preference, constraint, or convention worth remembering ("we always use pnpm", "never commit to main directly")
• A non-obvious fact about the code is established — an invariant, a workaround, a known gotcha
• The user explicitly says "remember this" or "save that"
• A decision is made that future sessions should know about

Use mode: "text" for standalone facts. Use mode: "messages" only when the full conversational turn is load-bearing.

What NOT to save

• Ephemeral task state (current WIP, what file you're editing)
• Facts already documented in CLAUDE.md, README, or recent git commits
• Anything the user can trivially re-derive from the code
• Debugging session scratch

When to package

Call memory_package when you need a curated, token-budgeted context block — e.g. for a fresh task where broad context matters more than a single fact lookup. Prefer memory_search for specific queries.

Scope

Scope flows automatically from the plugin config. Override per call only when the user explicitly asks to operate in a different scope ("check my personal memory", "only look at this repo").