lnm-brain-public: Instructions file for Claude Code

LNM Brain — Claude / Agent Memory Configuration v10.2.2

This is a reusable integration template. It must point to the current user's own private LNM Brain deployment; it must never contain another person's captured memory, credentials, or profile data.

Before substantive work

1. Call session_context() at session start.
2. For work involving existing projects, preferences, prior decisions, or history, retrieve relevant Second Brain context before answering or acting.
3. Prefer the cheapest useful retrieval path:

session_context
→ get_entity_facts / query_triples
→ keyword_search
→ recall_brain
→ query_second_brain / semantic_search
→ ask_second_brain

Stop once enough evidence is available.

Capture rules

Persist durable outcomes such as:

• decisions and reasons;
• project-state changes;
• important code/architecture conclusions;
• research findings;
• preferences and instructions;
• troubleshooting outcomes worth remembering;
• blockers and next actions.

Skip pure greetings, acknowledgements, duplicate captures, and low-value tool noise.

Use ingest_to_second_brain for the full pipeline. Use capture_to_second_brain as a lighter fallback where appropriate.

SESSION-HANDOFF

After a substantive session, write a compact handoff with:

Surface: <surface>
Time: <time>
Topic: <topic>

## State
<where things stand>

## Decisions Made
- <decision> → <reason>

## Avoided Paths
- <rejected path> → <reason>

## Actions Taken
- <completed work>

## Key Insights
- <durable learning>

## Key Data & Variables
- <versions, IDs, URLs, thresholds>

## Next Actions
1. <next action>

## Open Questions / Blockers
- <blocker>

## Confidence Flags
- Confirmed: <verified>
- Assumed: <not verified>

## Trail
<recent related handoffs>

The handoff should let a cold agent recover most of the working state quickly without loading raw history.

Runtime MCP interface

The v10.2.2 reference worker registers 27 MCP tools. The worker's tools/list response is authoritative: