memory · diff
git:20260610.be2a852 to git:20260610.c15894f
235 added, 4 removed. Audit A to A.
---
name: memory
description: "Knowledge system ops — recall, pollinate, audit docs. Subcommands: recall, pollinate, review. Use for pattern queries, cross-client transfer, or audits."
model: haiku
effort: medium
keywords: [knowledge, recall, patterns, cross-pollinate, audit, memory]
task_strategies: [investigation, spike]
stream_affinity: [research, tech-debt]
argument-hint: "[recall|pollinate|review|review --audit] [query]"
group: learning
allowed-tools:
- Agent
- AskUserQuestion
- Bash
- Glob
- Grep
- Read
status: stable
growth_stage: evergreen
---
+ # Memory
- <!-- PROCEDURE_FILE: procedures/memory.md -->
- This skill's full procedure is in a separate file for startup performance (ADR-034).
- Read and execute `../../procedures/memory.md` resolved against this skill's base directory (the path announced when the skill loads) — i.e. `{base-dir}/../../procedures/memory.md`. This form is valid in both the repo layout and the deployed-plugin layout.
- If the path doesn't resolve, use Glob to find `**/procedures/memory.md`.
+ Unified interface for the knowledge system. Replaces `/pattern-recall`, `/cross-pollinate`, and `/knowledge-review`.
+
+ ## Subcommand Routing
+
+ Parse `$ARGUMENTS` for the subcommand:
+
+ - `/brana:memory recall [query]` or `/brana:memory [query]` — search patterns (default)
+ - `/brana:memory pollinate [query]` — cross-client pattern transfer
+ - `/brana:memory review` — monthly knowledge health audit
+ - `/brana:memory review --audit [doc]` — cross-doc contradiction detection
+ - `/brana:memory audit` — surface lint+heal report and approve merges
+
+ If no subcommand recognized, default to **recall** with the full arguments as query. If no arguments at all, infer query from current project context.
+
+ ## Setup
+
+ ```bash
+ source "$HOME/.claude/scripts/cf-env.sh"
+ ```
+
+ ---
+
+ ## recall — Query Learned Patterns
+
+ 1. Use the query from `$ARGUMENTS` (after stripping the `recall` subcommand). If empty, infer from current project context (tech stack, current task, recent errors).
+
+ 2. **Primary path (ruflo available):**
+ Run `cd $HOME && $CF memory search --query "$QUERY"` to search the memory DB. Parse JSON values to extract `confidence`, `transferable`, and `recall_count` fields.
+
+ 3. **Fallback path (ruflo unavailable):**
+ Search `~/.claude/projects/*/memory/` for relevant MEMORY.md files. Grep for keywords from the query.
+
+ 3a. **Local per-pattern files** (fallback when ruflo unavailable):
+ Scan `~/.claude/projects/{project-hash}/memory/pattern_*.md`. For each file whose body
+ matches query keywords, surface it as a pattern result with its `confidence` frontmatter field.
+ These are local-only patterns not yet indexed in ruflo (git-durable, no cap).
+ Note: `~/.claude/memory/patterns.md` is retired — do not read it for new sessions.
+
+ 3b. **Local knowledge-staging.md** (always):
+ Read `~/.claude/memory/knowledge-staging.md`. For each `## slug` section matching the query, surface the claim and its `**Promote to:**` destination. Label clearly as "staging — not yet promoted."
+
+ 4. **Group results by confidence tier:**
+
+ ```
+ ## Proven patterns (confidence >= 0.7)
+ - [pattern] — confidence: X, recalls: N, source: PROJECT, transferable: yes/no
+
+ ## Quarantined patterns (confidence >= 0.2, < 0.7)
+ - [pattern] — confidence: X, recalls: N, source: PROJECT (treat with caution)
+
+ ## Suspect patterns (confidence < 0.2)
+ - [pattern] — confidence: X, recalls: N, source: PROJECT (previously demoted)
+ ```
+
+ 5. If no patterns found, say so explicitly — don't hallucinate past experience.
+
+ ---
+
+ ## pollinate — Cross-Client Pattern Transfer
+
+ 1. Detect current project's tech stack and problem domain.
+
+ 2. **Primary path:** Run `cd $HOME && $CF memory search --query "$QUERY"` to find patterns across all clients. Filter for **transferable patterns from other clients** only.
+
+ 3. **Fallback path:** Scan `~/.claude/projects/*/memory/MEMORY.md` from OTHER projects. Grep for technology and pattern type matches.
+
+ 4. Filter: only show patterns marked `transferable: true` or with confidence > 0.7.
+
+ 5. For each pattern show: source project, the pattern (problem + solution), confidence, why it might be relevant.
+
+ 6. Note: cross-pollinated patterns should be validated in the current project context before trusting them.
+
+ ---
+
+ ## review — Monthly Knowledge Health Audit
+
+ 1. **Gather stats** from ReasoningBank:
+ - Run `cd $HOME && $CF memory list --namespace pattern --limit 100`
+ - For each pattern (skip `test:*`), retrieve: `cd $HOME && $CF memory retrieve -k "KEY" --namespace pattern --format json`
+ - Parse: `confidence`, `transferable`, `recall_count`, `project`
+
+ 2. **Compute health metrics:**
+
+ ```
+ ## Knowledge Health Snapshot — YYYY-MM-DD
+
+ ### Overview
+ - Total patterns: N
+ - By project: project-a (N), project-b (N), ...
+
+ ### Confidence Distribution
+ - Proven (>= 0.7): N | Quarantined (0.2-0.7): N | Suspect (< 0.2): N
+
+ ### Transferability
+ - Transferable: N | Project-specific: N
+
+ ### Recall Activity
+ - Never recalled: N | 1-2 recalls: N | 3+ recalls (promote?): N
+
+ ### Staleness
+ - Stored > 60 days, never recalled: N (demotion candidates)
+ ```
+
+ 3. **Check local stores** (per-pattern files + knowledge-staging.md):
+ - Count `pattern_*.md` files in `~/.claude/projects/{project-hash}/memory/`. No cap. Surface count.
+ - Count `##` sections in `~/.claude/memory/knowledge-staging.md`. Cap: 30, warn-at: 20. Surface count + status.
+ - Add to health snapshot:
+ ```
+ ### Local Stores
+ - pattern_*.md files: N (no cap — per-pattern file model, ADR-039)
+ - knowledge-staging.md: N/30 entries (warn-at 20) — [OK | ⚠ Near cap | ✗ AT CAP]
+ ```
+ - If at or above warn-at: flag as action item — list staging entries with no `**Promoted:**` date.
+
+ 4. **Flag items:** promotion candidates (3+ recalls, still quarantined), staleness candidates, suspect patterns, staging entries past 30 days without promotion.
+
+ 5. **Suggest actions** — present options, let user decide. If no promotion/demotion/staleness candidates exist and all metrics are within thresholds (quarantine < 30%, staleness < 20%, proven > 50%), report "No action needed" with the numbers.
+
+ 6. **Backup** if changes made:
+ ```bash
+ "$HOME/.claude/scripts/backup-knowledge.sh"
+ ```
+
+ ---
+
+ ## review --audit — Cross-Doc Contradiction Detection
+
+ Traverses docs via formal `[doc NN](path)` links and flags factual contradictions. Works at the knowledge layer (doc vs doc), complementing `/brana:reconcile` (spec vs implementation).
+
+ ### Scope
+
+ - `/brana:memory review --audit` — audit all 5 reflections + both CLAUDE.md files (default)
+ - `/brana:memory review --audit [doc]` — audit a specific doc and everything it links to
+
+ ### Assertion Types to Extract
+
+ | Type | Pattern | Example |
+ |------|---------|---------|
+ | **Count** | `N (skills\|agents\|hooks\|rules\|docs\|dimensions\|reflections)` | "34 deployed skills" |
+ | **Version** | `v\d+\.\d+(\.\d+)?` or version-like strings | "v0.6.0", "all-MiniLM-L6-v2" |
+ | **Component list** | Markdown tables or bullet lists enumerating named items | Agent table, skill catalog |
+ | **Architecture claim** | Statements about what components do or how they connect | "claude-flow is the memory layer" |
+ | **Process claim** | Workflow descriptions with arrows or numbered sequences | "DDD → SDD → TDD", "dimension → reflection → roadmap" |
+
+ ### Steps
+
+ 1. **Select target docs.** Default: `reflections/08-diagnosis.md`, `reflections/14-mastermind-architecture.md`, `reflections/29-venture-management-reflection.md`, `reflections/31-assurance.md`, `reflections/32-lifecycle.md`, `.claude/CLAUDE.md`, `system/CLAUDE.md`. Or the single doc specified by user.
+
+ 2. **Extract assertions.** For each target doc, read it and extract factual claims:
+ - Grep for count patterns: `\b\d+\s+(skills?|agents?|hooks?|rules?|commands?|dimensions?|reflections?|patterns?|docs?)\b`
+ - Grep for version patterns: `\bv?\d+\.\d+(\.\d+)?\b` in non-URL contexts
+ - Identify component lists: tables with `|` separators listing named items
+ - Note architecture and process claims in prose
+
+ 3. **Verify counts against reality.** For verifiable counts:
+ - Skills: `ls system/skills/ | wc -l`
+ - Agents: count entries in `system/agents/`
+ - Rules: `ls system/rules/*.md | wc -l`
+ - Hooks: grep hook types in `system/hooks/` or settings.json
+ - Dimension docs: `ls docs/dimensions/*.md | wc -l` (via symlink)
+ - Reflection docs: `ls docs/reflections/*.md | wc -l`
+
+ 4. **Cross-reference assertions.** For each assertion, search other docs that mention the same topic (using formal links as the traversal graph). Flag when:
+ - Two docs state different counts for the same thing
+ - A version number in one doc doesn't match another
+ - A component list in doc A has items not in doc B's list (or vice versa)
+ - An architecture claim in one doc contradicts another
+
+ 5. **Report.** Output in errata-compatible format:
+
+ ```
+ ## Audit Report — YYYY-MM-DD
+
+ ### Contradictions Found: N
+
+ #### C-001: [title] — SEVERITY
+ - **Location:** [doc NN](path.md), line ~N
+ - **Claim:** "34 deployed skills"
+ - **Reality/Conflict:** actual count is 35 (verified via ls system/skills/)
+ - **Suggested fix:** update count to 35
+
+ #### C-002: [title] — SEVERITY
+ - **Location:** [doc NN](path.md) vs [doc MM](path.md)
+ - **Claim A:** "4 hook types"
+ - **Claim B:** "5 hooks: PreToolUse, SessionStart, SessionEnd, PostToolUse, PostToolUseFailure"
+ - **Suggested fix:** update doc NN to reflect 5 hook types
+
+ ### Verified Assertions: N
+ - Skills count: 35 ✓ (verified in 2 docs)
+ - Agent count: 10 ✓ (consistent across 2 docs)
+ ...
+ ```
+
+ 6. **Severity classification:**
+ - **HIGH**: count or version is wrong (leads to wrong implementation decisions)
+ - **MEDIUM**: component list is incomplete or stale (missing items)
+ - **LOW**: prose claim is imprecise but not misleading
+
+ 7. **Offer to apply fixes.** For HIGH/MEDIUM items, propose edits. Don't auto-apply — present and let user decide. If approved, apply as errata to [doc 24](24-roadmap-corrections.md).
+
+ ---
+
+ ---
+
+ ## audit — Surface Lint+Heal Report
+
+ 1. **Check report exists and is fresh.**
+ - Read `~/.claude/lint-heal-report.md`
+ - If missing: tell the user the report hasn't been generated yet. Suggest: `./system/scripts/lint-heal.sh --dry-run`
+ - If older than 7 days: warn that findings may be stale, then continue.
+
+ 2. **Surface the summary.**
+ - Read the `## Summary` section of the report (first 30 lines are usually enough).
+ - Show: total candidates found, breakdown by category (duplicates, contradictions, frontmatter gaps, concept refs, pattern_*.md near-duplicate slugs, knowledge-staging.md cap status).
+
+ 3. **Interactive approve-merges flow** (for duplicate and contradiction candidates only):
+ - List each HIGH/MEDIUM candidate with: source file(s), the conflict or duplication, proposed action (archive/merge/update).
+ - For each: ask user to approve, skip, or edit the proposed action using AskUserQuestion with options `["approve (Recommended)", "skip", "edit action"]`.
+ - Only apply approved fixes. Don't touch skipped items.
+
+ 4. **Apply approved fixes:**
+ - Archive: move file to `~/.claude/memory/archive/YYYY-MM-DD/` (create dir if needed).
+ - Merge: prompt user for the canonical version, then update the surviving file and archive the duplicate.
+ - Update frontmatter: write missing `name:`, `description:`, `type:` fields inferred from filename and content.
+
+ 5. **Write a completion summary** showing how many items were resolved vs skipped.
+
+ ---
+
+ ## Rules
+
+ - **Don't auto-modify patterns.** Review reports and suggests. The user decides.
+ - **Skip test data.** Entries with keys starting with `test:*` are from the test suite.
+ - **Ask for clarification whenever you need it.** If the query is too broad, context unclear, or results ambiguous — ask.