recall

Intent

Execute recall to retrieve relevant prior decisions, rationale, and patterns from memory to maintain consistency and learn from past interactions.

Success criteria:

• Relevant memories are retrieved based on query
• Prior decisions and their rationale are surfaced
• Consistency with past reasoning is maintained
• Retrieved information is grounded with timestamps and context

Compatible schemas:

• schemas/output_schema.yaml

Inputs

Inputs not listed here are outside the capability's contract: a workflow step that binds them gets W604 (STANDARD v1.1 §5.5).

Procedure

1. Parse query intent: Understand what the key asks to recall
   • Identify if it is about a decision, pattern, fact, or context
   • Extract key terms for memory search
   • Determine if exact match or semantic similarity needed
2. Scope memory search: Identify which memory stores to query
   • session: Current conversation context
   • project: CLAUDE.md, knowledge files, local docs
   • global: All accessible memory stores
3. Search memory stores: Query relevant sources
   • CLAUDE.md for project-level decisions and patterns
   • Session context for recent interactions
   • Knowledge files for domain-specific learnings
   • Use Grep for exact term matches, Read for context
4. Rank by relevance: Score and filter results
   • Relevance to query terms
   • Recency (more recent = higher weight unless historical needed)
   • Authority (explicit decisions > implicit patterns)
   • Filter out matches below a similarity threshold
5. Extract rationale: For each recalled item, capture reasoning
   • Original decision context
   • Factors that influenced the decision
   • Constraints or assumptions at the time
6. Check consistency: Verify recalled items are internally consistent
   • Flag contradictions between recalled items
   • Note if context has changed since original decision
7. Format output: Structure results according to output contract

Output Contract

Return a structured object:

found: boolean  # Whether anything matched the key; when false, data is the `default` input
stored_at: string | null  # ISO 8601 date-time the top match was recorded
data:  # Recalled items, best match first
  - id: string  # Unique identifier for this memory
    type: decision | pattern | fact | context
    content: string  # The recalled information
    rationale: string | null  # Why this decision was made
    source: string  # Where this was found (CLAUDE.md:42, session, etc.)
    timestamp: string | null  # When originally recorded
    relevance_score: number  # 0.0-1.0 match to query
    context:
      project: string | null
      conversation_id: string | null
      original_query: string | null
query_interpretation: string  # How the query was understood
search_scope:
  stores_searched: array[string]
  time_range_applied: object | null
total_matches: integer  # Total before filtering
returned_matches: integer  # After relevance filtering
consistency_check:
  consistent: boolean
  contradictions: array[object] | null
confidence: number  # 0.0-1.0 based on match quality
evidence_anchors: array[string]  # Sources for recalled items
assumptions: array[string]  # Assumptions about memory validity

Field Definitions

Examples

Example 1: Recalling a Project Decision

Input:

key: "error handling approach"

Output:

found: true
stored_at: "2024-01-10T09:15:00Z"
data:
  - id: "decision_err_001"
    type: decision
    content: "Use Result<T, Error> pattern for all fallible operations; avoid exceptions"
    rationale: "Explicit error handling improves debuggability and makes failure paths visible in types"
    source: "CLAUDE.md:127"
    timestamp: "2024-01-10T09:15:00Z"
    relevance_score: 0.95
    context:
      project: "api-service"
      conversation_id: "conv_abc123"
      original_query: "How should we handle errors in this codebase?"
  - id: "pattern_err_001"
    type: pattern
    content: "Wrap external API errors with domain-specific error types before propagating"
    rationale: "Prevents leaking implementation details; enables consistent error messages"
    source: "CLAUDE.md:134"
    timestamp: "2024-01-10T09:20:00Z"
    relevance_score: 0.88
    context:
      project: "api-service"
      conversation_id: "conv_abc123"
      original_query: null
query_interpretation: "Looking for decisions and patterns related to error handling in the codebase"
search_scope:
  stores_searched: ["CLAUDE.md", "docs/architecture.md"]
  time_range_applied: null
total_matches: 5
returned_matches: 2
consistency_check:
  consistent: true
  contradictions: null
confidence: 0.92
evidence_anchors:
  - "CLAUDE.md:127"
  - "CLAUDE.md:134"
assumptions:
  - "CLAUDE.md reflects current project decisions"
  - "No conflicting decisions recorded elsewhere"

Evidence pattern: Each recalled item traced to specific file:line; rationale extracted from surrounding context.

Example 2: Recalling Session Context

Input:

key: "user's preferred output format"

Search session memory only, from 2024-01-16 onward.

Output:

found: true
stored_at: "2024-01-16T10:30:00Z"
data:
  - id: "ctx_fmt_001"
    type: context
    content: "User prefers YAML output over JSON for configuration files"
    rationale: "User stated 'I find YAML more readable' in earlier message"
    source: "session:message_42"
    timestamp: "2024-01-16T10:30:00Z"
    relevance_score: 0.97
    context:
      project: null
      conversation_id: "current_session"
      original_query: "Can you output this as YAML instead?"
query_interpretation: "Looking for user preferences about output formatting from current session"
search_scope:
  stores_searched: ["session"]
  time_range_applied:
    after: "2024-01-16T00:00:00Z"
    before: null
total_matches: 1
returned_matches: 1
consistency_check:
  consistent: true
  contradictions: null
confidence: 0.97
evidence_anchors:
  - "session:message_42"
assumptions:
  - "User preference still valid within session"
  - "No subsequent preference change"

Example 3: Detecting Contradictory Memories

Input:

key: "database choice"

Output:

found: true
stored_at: "2024-01-05T14:00:00Z"
data:
  - id: "decision_db_001"
    type: decision
    content: "Use PostgreSQL for all persistent storage"
    rationale: "Strong ACID guarantees, mature ecosystem"
    source: "CLAUDE.md:45"
    timestamp: "2024-01-05T14:00:00Z"
    relevance_score: 0.94
    context:
      project: "backend-service"
      conversation_id: "conv_xyz789"
      original_query: "Which database should we use?"
  - id: "decision_db_002"
    type: decision
    content: "Use SQLite for local development and testing"
    rationale: "Faster setup, no external dependencies"
    source: "docs/dev-setup.md:23"
    timestamp: "2024-01-12T11:30:00Z"
    relevance_score: 0.85
    context:
      project: "backend-service"
      conversation_id: "conv_def456"
      original_query: "How to speed up local dev?"
query_interpretation: "Looking for decisions about database selection"
search_scope:
  stores_searched: ["CLAUDE.md", "docs/"]
  time_range_applied: null
total_matches: 2
returned_matches: 2
consistency_check:
  consistent: false
  contradictions:
    - items: ["decision_db_001", "decision_db_002"]
      attribute: "database"
      resolution_hint: "Different contexts: production vs development"
confidence: 0.75
evidence_anchors:
  - "CLAUDE.md:45"
  - "docs/dev-setup.md:23"
assumptions:
  - "Both decisions are currently valid"
  - "Context difference (prod vs dev) resolves contradiction"
next_actions:
  - "Clarify if query refers to production or development"
  - "Document environment-specific database strategy"

Verification

• Query intent correctly interpreted
• All relevant memory stores searched within scope
• relevance_score reflects actual match quality
• Contradictions detected and flagged
• evidence_anchors provide traceable source references

Verification tools: Read (to validate memory file contents), Grep (to search across memory stores)

Safety Constraints

• mutation: false
• requires_checkpoint: false
• requires_approval: false
• risk: low

Capability-specific rules:

• Do not recall or expose credentials, secrets, or PII from memory
• Respect memory scope boundaries; do not access global when session requested
• Flag stale memories (>30 days) with lower confidence
• Do not fabricate memories; if nothing is found return found: false with data set to the default input (or null)

Composition Patterns

Commonly follows:

• receive - After receiving a request, recall relevant context
• search - Search results may trigger memory lookup for similar past queries

Commonly precedes:

• decide - Recalled decisions inform new decision-making
• plan - Past patterns guide new plan creation
• explain - Recalled rationale used to explain current approach
• persist - New decisions should be consistent with recalled ones

Anti-patterns:

• Never override recalled decisions without explicit user request
• Never recall without checking for contradictions
• Avoid recalling from stale sources without flagging age

Workflow references:

• Memory recall is implicit in many workflows as context loading
• Pairs with persist for the remember/recall cycle