| name | muninn-memory-protocol |
| description | Use this skill when a client should adopt the shared Muninn memory workflow for continuity across sessions. Trigger when setting up Muninn-backed recall/write-back habits, wiring a client to the shared helper, or standardizing memory behavior across multiple cognitive clients. |
Muninn Memory Protocol
Use this skill when you need a client to participate in the shared Muninn memory experiment.
Read MUNINN_CLIENT_MEMORY_PROTOCOL.md first for the portable contract.
Purpose
This skill standardizes how a client should:
- retrieve before meaningful work
- write back after durable outcomes
- keep memory atomic and fragmented into short, meaningful bursts
- rely on the shared helper instead of hand-rolling MCP ceremony
Ownership Split
Muninn is the continuity layer across clients. It is not the source of truth, the task tracker, or a transcript archive.
Use the right owner:
- repo docs and code store implemented truth
- Intel Graph stores structure, seams, work coordination, decisions, and verification evidence
docs/task.md stores active execution work
- Muninn stores why something matters next time: compact decisions, learned preferences, reality gaps, validation outcomes, and continuity handles
Use The Shared Helper
Preferred transport helper:
Do not reimplement the MCP handshake ad hoc unless you are intentionally porting the helper into another runtime.
Default Habit
Before meaningful work:
- run
bootstrap
- run
where-left-off
- run
recall
After durable outcomes:
- run
remember for atomic facts, preferences, and outcomes in short, meaningful bursts
- run
decide for explicit decisions with rationale
Keep writes small:
remember: 1-3 sentences, preferably under ~300 chars, hard ceiling ~500
decide: concise rationale, preferably under ~500 chars, hard ceiling ~800
Use a small tag vocabulary only when it helps retrieval:
flesh-out
decision
reality-gap
validation
follow-up
operator-preference
At closeout, prefer a memory delta:
Memory delta:
- Decision:
- Reality gap:
- Validation:
- Next seam:
- Operator preference:
Each filled line should usually become one atomic write. Omit empty lines. Do not store transcripts, noisy logs, routine task churn, or summaries of docs already committed in the repo.
Failure Rule
The helper must be allowed to block.
If python3 scripts/muninn_mcp.py bootstrap fails:
- alert the user/operator immediately
- do not pretend retrieval already happened
- do not continue with meaningful work until explicit approval is given to proceed without Muninn
The protocol is continuity first, not silent degradation dressed up as professionalism.
Truth Rule
Recalled memory is guidance, not authority.
If recalled memory conflicts with current observed repo/runtime truth:
- trust observed truth
- note the mismatch if it matters
Why The Helper Is Shared Instead Of Skill-Only
This skill is a client adapter.
The helper is transport infrastructure.
If the helper exists only inside a client-specific skill:
- other cognitive clients cannot reuse it cleanly
- the protocol becomes trapped in one client packaging format
- every new client has to rediscover the same handshake logic
So the canonical helper should live in normal versioned project infrastructure, and client skills should wrap it.