Grilling: what does continuous procedural-memory organization (skill mining) mean?

[antejavor]

Part of #259

Question

skills-graph today only stores/retrieves explicitly-defined skills (e.g. from a SKILL.md read, or an explicit add_skill call) — there is no mechanism that mines a new Skill out of observed repeated behavior. What should "continuous procedural-memory organization" mean concretely: what pattern in captured Action history should trigger a mined Skill, what's the threshold, and does a mined Skill get written automatically or require human/agent approval before becoming real? Resolve with the driving user via /grilling + /domain-modeling.

[antejavor]

Resolution

Procedural-memory organization resolves via /grilling + /domain-modeling with the driving user:

Two separate mining processes, not a unified pass — they detect fundamentally different kinds of repetition and conflating them risked a design mediocre at both:

• Sequence-mining: over actions-graph's chronological Action/FOLLOWED_BY data — repeated tool-call sequences across sessions. Pure graph pattern-matching, no LLM required.
• Content-mining: over files touched/read during sessions — repeated patterns in file content (not just tool-call shape). Reuses the same unstructured2graph-style LLM-backed pipeline sessions-graph's reconciliation already has, pointed at a different corpus (touched files, not session transcripts).

Output shape: both processes write to a new node type, Procedure — deliberately distinct from Skill. Skill stays locked to the Agent Skills specification (agentskills.io), untouched by this design; a Procedure is not required to conform to it. Naming considered and rejected: MinedSkill (ties naming to Skill despite deliberately not being one), LearnedPattern (reads as the evidence, not the distilled output), ProceduralMemory (redundant with the memory-type name it's an instance of — every other memory type uses a concrete noun: Memory, Session, not SemanticMemory/EpisodicMemory).

Write policy: auto-write, no approval gate. Procedure's separateness from Skill is itself the safety mechanism — nothing pollutes the interoperable Skill surface just by a Procedure existing.

Explicitly out of scope: promotion from a mined Procedure into a real curated Skill — future work, not designed here. Also out of scope: the exact trigger/cadence of when mining runs — deferred to #263, which owns that question for all organizing jobs.