wiki-learn

Wiki learn — Memory

Persist patterns, preferences, pitfalls, and next steps across sessions in a single human-reviewable file: llm-wiki/.agent-memory.md. Plain Markdown (no DB). Optional: integrate with wiki-retro (recommendations → Next Steps).

Privacy: This file is assistant-facing notes, not canonical wiki/ evidence — but if the vault is in git, treat it like any tracked doc: do not record secrets, API keys, or confidential tokens here; use env or gitignored settings instead.

Pre-flight

1. Ensure llm-wiki/ exists. If .agent-memory.md is missing, create it from the template below (or append sections to an existing file).

Initial file template

# Agent memory (llm-wiki)

Optional cross-session notes. Edit freely. This file is **not** evidence for claims in `wiki/` — use `raw/` + `wiki/` for that.

## Research Patterns

## Preferences

## Pitfalls

## Next Steps

Operations (from $ARGUMENTS or user message)

Mode	Action
learn	Append a dated bullet under the right section: YYYY-MM-DD — <note> (context: …)
recall	Read the file and surface entries relevant to the current task or query
review	Show all sections; offer to prune, merge, or move items to wiki/log.md if they become project decisions

Steps

Step 1 — Classify the note

• Research Patterns — what source combos work, domains to avoid, query tricks
• Preferences — tone, page structure, PDF adapter choice, automation level
• Pitfalls — repeated failures (rate limits, paywalls, bad extractions)
• Next Steps — follow-up topics (often fed by wiki-retro)

Step 2 — Read or write

• learn: append; do not duplicate identical bullets unless reinforcing
• recall: quote 3–7 most relevant lines for the user’s question
• review: full file, then ask what to delete or rewrite

Step 3 — Integration

• wiki-research pre-flight: optionally recall Pitfalls + Preferences before classifying the request (see wiki-research — add one line “Read .agent-memory.md if present”).
• wiki-retro: may append Recommendations to Next Steps here.

Done looks like

• learn: new bullet(s) visible under the correct section with date.
• recall: user sees relevant memory without dumping the whole file unless review.
• review: user confirms edits or leaves file as-is.

Artifacts

Reads	Writes
llm-wiki/.agent-memory.md	llm-wiki/.agent-memory.md

Related skills

• wiki-research — consumes recall at session start (optional)
• wiki-retro — feeds Next Steps
• wiki-pipeline — may read memory before long runs
• wiki-session-memory — per-chat raw/memory/<session-id>.md; promote recurring patterns from session files into .agent-memory.md when useful.

Smoke check

• CLI: llm-wiki validate from the vault root (vault must exist).
• Prompt: Invoke this skill; confirm learn/recall targets llm-wiki/.agent-memory.md when relevant.

Optional: persona.name in config (default Gennie).