git:20260610.c15894f to git:20260813.b1e2ac9

---

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

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.

```bash

```

---

## 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.