ww-recall

name	ww-recall
description	Recall memories from World Weaver using multi-strategy retrieval
version	1.0.0
allowed-tools	["Bash","Read"]

WW Recall Skill

Multi-strategy memory retrieval from World Weaver's tripartite memory system.

Purpose

Search and retrieve relevant memories from:

1. Episodic Memory: Past experiences and events
2. Semantic Memory: Knowledge graph entities and relationships
3. Procedural Memory: Skills and how-to patterns

When to Use

Invoke this skill when:

• User asks "what did we do before?", "do you remember?"
• Need context from previous sessions
• Looking for relevant past experiences
• Searching for applicable skills
• Building context for a new task

MCP Tools Available

mcp__ww-memory__recall_episodes    - Search episodic memory
mcp__ww-memory__semantic_recall    - Search knowledge graph
mcp__ww-memory__spread_activation  - Graph traversal from seed
mcp__ww-memory__recall_skill       - Find applicable skills
mcp__ww-memory__query_at_time      - Point-in-time queries

Retrieval Strategies

Strategy 1: Semantic Search (Default)

Best for: General queries, finding similar content

mcp__ww-memory__recall_episodes(
  query: "your search query",
  limit: 10,
  session_filter: null  # All sessions, or specific session_id
)

Returns episodes ranked by:

• Semantic similarity to query (40%)
• Recency decay (25%)
• Outcome weight (20%)
• Importance/valence (15%)

Strategy 2: Temporal Search

Best for: "What did we do last week?", time-bounded queries

mcp__ww-memory__recall_episodes(
  query: "project work",
  limit: 10,
  time_filter: {
    after: "2025-11-20T00:00:00",
    before: "2025-11-27T23:59:59"
  }
)

Strategy 3: Knowledge Graph Search

Best for: Finding entities and their relationships

mcp__ww-memory__semantic_recall(
  query: "concept or entity name",
  limit: 10,
  include_relationships: true
)

Strategy 4: Spread Activation

Best for: Exploring connections from a known starting point

mcp__ww-memory__spread_activation(
  seed_entities: ["Entity A", "Entity B"],
  max_depth: 2,
  min_weight: 0.3
)

Returns entities connected to seeds, weighted by relationship strength.

Strategy 5: Skill Matching

Best for: "How do I...?", finding applicable procedures

mcp__ww-memory__recall_skill(
  query: "how to run tests",
  limit: 5,
  check_preconditions: true,
  context: {
    project: "ww",
    working_directory: "/home/aaron/ww"
  }
)

Strategy 6: Point-in-Time Query

Best for: "What did we know at time X?"

mcp__ww-memory__query_at_time(
  query: "project status",
  as_of: "2025-11-15T12:00:00"
)

Multi-Strategy Retrieval Workflow

For comprehensive recall, combine strategies:

1. Analyze Query

Determine what the user is looking for:

• Past events → Episodic (Strategy 1, 2)
• Concepts/entities → Semantic (Strategy 3, 4)
• How-to procedures → Procedural (Strategy 5)
• Historical state → Temporal (Strategy 6)

2. Execute Searches

Run appropriate strategies in parallel:

episodes = mcp__ww-memory__recall_episodes(query, limit=5)
entities = mcp__ww-memory__semantic_recall(query, limit=5)
skills = mcp__ww-memory__recall_skill(query, limit=3)

3. Merge Results

Combine results, removing duplicates by ID.

4. Rank Results

Re-rank by overall relevance:

• Query similarity
• Recency
• Source diversity (mix of episode/entity/skill)

5. Format Output

Present as structured context:

## Memory Context

### Recent Episodes (3 found)
1. **[2025-11-27]** Fixed batch query bug in Neo4j store
   - Outcome: success
   - Relevance: 0.92

2. **[2025-11-26]** Implemented session isolation tests
   - Outcome: success
   - Relevance: 0.85

### Related Knowledge (2 entities)
1. **Neo4j** (CONCEPT)
   - Graph database used for relationships
   - Connected to: Qdrant, Cypher, World Weaver

2. **Batch Queries** (CONCEPT)
   - Optimization pattern for N+1 elimination

### Applicable Skills (1 found)
1. **run-ww-tests**
   - Run World Weaver test suite
   - Preconditions: In WW directory, venv exists

Retrieval Parameters

Common Parameters

Parameter	Type	Description
query	string	Search query text
limit	int	Max results (default: 10)
session_filter	string	Filter by session ID
time_filter	object	After/before timestamps

Episode-Specific

Parameter	Type	Description
outcome_filter	string	success/failure/partial/neutral
min_valence	float	Minimum importance (0-1)

Entity-Specific

Parameter	Type	Description
entity_type	string	CONCEPT/PERSON/PLACE/etc
include_relationships	bool	Include connected entities

Skill-Specific

Parameter	Type	Description
check_preconditions	bool	Verify preconditions met
min_success_rate	float	Minimum skill success rate

Examples

Example 1: General Recall

User: "What have we worked on recently?"

Action:
mcp__ww-memory__recall_episodes(
  query="recent work projects tasks",
  limit=10,
  time_filter={after: "7 days ago"}
)

Output:
Found 8 recent episodes:
1. [Nov 27] Plugin architecture planning for World Weaver
2. [Nov 27] Fixed batch query bugs, session isolation
3. [Nov 26] Implemented HDBSCAN memory clustering
...

Example 2: Skill Lookup

User: "How do I run the integration tests?"

Action:
mcp__ww-memory__recall_skill(
  query="run integration tests",
  limit=3,
  check_preconditions=true
)

Output:
Found skill: run-ww-tests
Steps:
1. source .venv/bin/activate
2. pytest tests/integration/ -v -m integration
3. Check output for failures

Example 3: Entity Exploration

User: "What do we know about Neo4j?"

Action:
1. mcp__ww-memory__semantic_recall(query="Neo4j", limit=1)
2. mcp__ww-memory__spread_activation(seed_entities=["Neo4j"], max_depth=2)

Output:
## Neo4j (CONCEPT)
Graph database for relationship storage.

### Connected Entities:
- Cypher (query language) - strength: 0.9
- Qdrant (co-storage) - strength: 0.8
- World Weaver (uses) - strength: 0.95
- Batch Queries (optimization) - strength: 0.7

Example 4: Time-Bounded Search

User: "What did we do last week on the PhD project?"

Action:
mcp__ww-memory__recall_episodes(
  query="PhD dissertation xai",
  limit=10,
  time_filter={
    after: "2025-11-20",
    before: "2025-11-27"
  }
)

Output:
Found 4 episodes from Nov 20-27:
1. [Nov 25] Committee meeting preparation
2. [Nov 23] VGGFace2 dataset removal analysis
3. [Nov 21] Article C IEEE T-BIOM submission
4. [Nov 20] Experiment 6.1 planning

Quality Guidelines

When presenting recall results:

• Show relevance scores when available
• Include timestamps for temporal context
• Link related entities when relevant
• Highlight applicable skills
• Indicate result source (episodic/semantic/procedural)

Error Handling

If recall fails:

1. Check MCP server connectivity
2. Try simpler query (fewer filters)
3. Fall back to keyword-based grep of local files
4. Inform user what memories are unavailable