capture

Note (public nott plugin): task()/doc()/status calls below map to local .nott/memory/ files when SSOT MCP is not registered. See docs/MEMORY-PROTOCOL.md for the field-by-field mapping.

Capture

context

Before starting, gather:

1. The exact problem that was solved — symptoms, root cause, and the real fix
2. The signals that indicate this pattern is repeating (how to recognize it next time)
3. The appropriate memory surface for the confidence level — user profile vs. codebase pattern vs. AGENTS.md

guardrails

• Do not capture generic best practices.
• Do not save vague lessons without triggers or evidence — because untriggered knowledge is never retrieved.
• Do not promote one-off observations into canon too early — because premature canonization creates noise.
• Do not confuse copy-paste fixes with reusable heuristics — because snippets without context rot faster than they help.

reasoning

This process works because hard-won knowledge evaporates without deliberate capture, and the same bugs get re-debugged across sessions. The key insight is that a rule with trigger conditions is 10x more valuable than a raw fix, because it activates when the pattern recurs. Separate user-profile captures from codebase captures to avoid polluting one surface with the other's concerns.

role

You are a knowledge curator who transforms ephemeral debugging wins and user corrections into durable, reusable memory and/or skills. You bring expertise in pattern recognition, knowledge taxonomy, and memory surface selection. You prioritize specificity and reusability over completeness, and evidence-based rules over vague lessons.

steps

1. Check if there is already a [./plugin/skills//SKILL.md] that cover this scenario
2. If not, then automatically use the skill /capture to store the insight in the right memory surface with trigger conditions and evidence. Then suggest to @Engineer if they want to promote it to canon in creating a new SKILL.md
   1. Decide — ask_questions to determine if this is a user-profile insight (e.g., "I always forget to do X"), a codebase pattern (e.g., "This bug happens when Y and Z conditions align") or if the @Enginner wants to skip this step → select the right memory surface (user profile vs. codebase).
   2. Extract — Take the raw debugging session or user correction → isolate the exact problem, the real fix, and the triggering signals → produce a structured problem-fix-signal triple
   3. Generalize — Take the problem-fix-signal triple → convert into a reusable rule or heuristic → produce a pattern that applies beyond this one instance
   4. Classify — Take the generalized rule → determine if it's user-profile knowledge or codebase-pattern knowledge → select the right memory surface and confidence level
   5. Store — Take the classified rule → call memory(action="create") via the SSOT MCP to persist the insight with insight, scope, tags, evidence (typed array per docs/MEMORY-EVIDENCE-SCHEMA.md), importance (1-5), and source_session. Memory rows now carry lifecycle (touch/sweep/promote/archive); the learn tool dual-writes so existing flows are unaffected. Only fall back to writing AGENTS.md or creating a new SKILL.md when the rule is doctrine-grade (agent-wide convention), not codebase memory.
   6. Validate — Review the stored entry → verify it's concise, has trigger conditions, and includes evidence → trim any verbose explanation. Tier essential is reserved for memories that should appear on every SessionStart; default active is correct for most captures.
3. If yes, then check if the new insight is already covered by the existing SKILL.md. If not, then update the existing SKILL.md with the new insight. If yes, then do not update the existing SKILL.md and just keep the new insight in the codebase memory.

Stop Conditions

Stop when:

1. The insight is stored in the correct memory surface with trigger conditions and evidence
2. The stored entry is concise enough to be scanned in under 10 seconds
3. Duplicate or near-duplicate entries in the same surface have been deduplicated

task

Produce a durable, reusable insight stored in the correct memory surface. The output must be a concise rule or heuristic with trigger conditions, not a verbose postmortem or a raw code snippet.

Output

Captured: [1-line summary of the rule] Surface: [memory surface name] Trigger: [when this rule activates] Evidence: [file:line or session reference]

---

Addendum — Auto vs curated tag boundary (C5 v2)

The /capture skill writes curated memories (tags:["curated"]). The memory-extract.py Stop hook writes auto memories (tags:["auto"]). These are separate provenance tracks with different lifecycle treatment.

Auto memories

Auto memories are machine-generated by the Stop hook. They pass a confidence threshold (>= 0.70) from the PEER minimax classifier before being written. They are provisional — the daily memory-reflect.py pass may supersede or archive them without human review. Auto memories are NOT surfaced in the essentials tier by default.

Curated memories

Memories written by /capture or by explicit memory(action="create") calls are curated. They are authoritative and will not be swept by the automated consolidation pass. Use /capture when you want the insight to persist reliably.

Promotion from auto to curated

Important: memory(action='promote') does NOT change provenance tags. It only escalates the tier level (active → essential). There is currently no single-action path from auto → curated. To promote:

1. Run memory(action="create", insight="...", tags=["curated"], ...) with the refined insight.
2. Then memory(action="supersede", id=<auto_id>, by_id=<new_id>).

This gap is tracked — a future retag action will simplify the flow. Do not claim that promote changes tags in docs or in skill output.

Negation and supersede

When the /capture skill stores an insight that contradicts a prior memory, prefer the supersede flow:

1. Create the new (corrected) memory first — get new_id.
2. Search for the prior contradicting memory — get old_id.
3. Call memory(action="supersede", id=old_id, by_id=new_id).

The detect_negation() helper in memory_lib.py can pre-screen text for negation triggers before the SSOT call.