rc-memory

name	rc-memory
description	The single place for RC's durable, cross-session memory and learning. Records and recalls curated project facts (decisions, conventions, gotchas, glossary) AND distills recurring corrections/workflows into confidence-scored learnings — all as plain markdown under .rc/memory/. Use to consult memory before working and to record durable facts or distilled learnings afterward. Do not use for workflow/task-scoped notes (use rc-workflow-memory) or for on-demand reflection on a diff (use lesson-learned).
user-invocable	true
model	sonnet
effort	medium

Memory

The one place for what RC should remember about this project across turns and sessions. It is project-scoped only — never copy memory or learnings to another project or a global location. It holds two kinds of durable knowledge, both as plain markdown under .rc/memory/ — no database, no binary, and this skill never edits source code:

1. Facts — curated decisions, conventions, gotchas, glossary, and context.
2. Learnings — recurring corrections/workflows distilled into confidence-scored trigger → action rules (the continuous-learning loop fed by the observe hook).

It is separate from rc-workflow-memory (ephemeral, per-workflow notes under .rc/tasks/<slug>/) and from lesson-learned (on-demand reflection on a diff, not a store).

Store layout

.rc/memory/
  INDEX.md              # one line per fact — the recall entry point
  <slug>.md             # one durable fact per file
  LEARNINGS.md          # distilled trigger→action learnings, confidence-scored
  observations.jsonl    # raw capture appended by the observe hook (transient)

Resolve the project's .rc base directory (nearest .rc walking up; default ./.rc).

A — Facts

Each fact is one <slug>.md with frontmatter + a short body:

---
title: <short title>
scope: decision | convention | gotcha | glossary | context
key: <stable-kebab-slug or omit>
tags: [a, b]
source: <skill/command that produced it>
created: <YYYY-MM-DD>
updated: <YYYY-MM-DD>
---

<the fact in a few sentences. For a decision, add **Why:** and, if useful, **How to apply:**.>

INDEX.md holds one line per fact, newest last: - [<title>](<slug>.md) — <scope> — <one-line hook>. It is the only file you read to decide relevance; never dump fact bodies into it.

Consult (before implementing/deciding/planning): read INDEX.md, scan the one-line hooks for the task's terms, open only the matching <slug>.md. For a large store, grep -ri "<term>" .rc/memory/.

Record a fact only when all three hold: (1) a future run would need it to avoid a mistake or rediscovery; (2) it is durable for this project, not one task's execution; (3) it is NOT already obvious from the repo, git history, PRD, or techspec. Reuse the same slug/key to update instead of duplicating. Keep each fact short; never store secrets, tokens, credentials, large code blocks, stack traces, or raw session logs.

B — Learnings (the continuous-learning loop)

When the same correction, error resolution, or workflow recurs, stop re-learning it each session and make it a learning: one trigger → action rule with a confidence score and evidence. Capture is the observe hook (on by default; RC_INSTINCTS=0 to opt out), appending {ts, tool, target} lines to .rc/memory/observations.jsonl. This skill is the distill/curate side, run deliberately. The memory-load SessionStart hook surfaces existing facts/learnings at the start of each session and nudges you to run this skill once observations accumulate (≥40 lines).

LEARNINGS.md groups atomic learnings by domain; each is one line:

## code-style
- [0.7] when adding a new error path → wrap with the project's error-wrap helper  (evidence: corrected 3×; updated 2026-07-07; seen 4)

• Confidence 0.3–0.9; start fresh at 0.4–0.5. One trigger → one action (split if you need "and").
• Evidence is why it exists — no evidence, not a learning.

Distill: (1) read observations.jsonl (if present) + the current session — corrections, resolved errors, sequences repeated 3+×, stated preferences. (2) Cluster into atomic trigger → action candidates; discard one-off or already-obvious ones. (3) Merge with LEARNINGS.md: new → add at 0.4–0.5 with seen 1; recurred uncontradicted → raise confidence (+0.1–0.2, cap 0.9), bump seen, refresh date; contradicted this session → lower (−0.2–0.3), drop below 0.3; stale + low → drop. (4) Write LEARNINGS.md sorted by domain then confidence, then truncate observations.jsonl so the next pass starts clean. (5) Report what was added/reinforced/weakened/dropped.

Critical rules

• When a fact or learning is contradicted by the repo or the user, correct or delete it — trust the repo over stale memory.