Agentic memory

Agentic Memory Skill

Ontological memory for Alhazen: the agent understands the schema, composes queries dynamically, and navigates the full knowledge graph across all skill namespaces.

Core capability: Three-stage retrieval — plan queries from schema understanding, execute multi-step graph + vector searches, organize results with provenance.

Write-side entities:

• operator-user — person running Alhazen; 10-domain personal context
• memory-claim-note — crystallized proposition (fact-type: knowledge | decision | goal | preference | schema-gap)
• episode — process account of a work session with operation tracking (what was created/modified and why)
• entity-alias — cross-skill resolution (same real-world entity as multiple TypeDB entities)

Quick Start

# Understand the schema
uv run python .claude/skills/agentic-memory/agentic_memory.py describe-schema 2>/dev/null

# Run any TypeQL query
uv run python .claude/skills/agentic-memory/agentic_memory.py query \
  --typeql 'match $p isa person, has name $n; $n contains "Burns"; fetch { "id": $p.id, "name": $n };' 2>/dev/null

# Semantic search across all embedded collections
uv run python .claude/skills/agentic-memory/agentic_memory.py search \
  --query "CRISPR delivery mechanisms" --limit 5 2>/dev/null

# Create operator profile
uv run python .claude/skills/agentic-memory/agentic_memory.py create-operator \
  --name "Gully Burns" --identity "Principal AI Architect" --role "Building knowledge graphs" 2>/dev/null

# Consolidate a fact into long-term memory
uv run python .claude/skills/agentic-memory/agentic_memory.py consolidate \
  --content "TypeDB 3.x sub! operator returns direct parent only" \
  --subject <entity-id> --fact-type knowledge --confidence 0.9 2>/dev/null

# Create episode with operation tracking
uv run python .claude/skills/agentic-memory/agentic_memory.py create-episode \
  --skill agentic-memory --summary "Fixed schema hierarchy" 2>/dev/null
uv run python .claude/skills/agentic-memory/agentic_memory.py link-episode \
  --episode <ep-id> --entities <id1>,<id2> \
  --operation-type created --rationale "Created during schema fix session" 2>/dev/null

Three-Stage Retrieval Pipeline

When answering questions about the knowledge graph, follow this three-stage pattern:

Stage A: Plan Retrieval

1. Understand the schema: Call describe-schema to see what entity types, relations, and embedding indexes exist
2. Identify relevant types: Map the user's question to TypeDB types (e.g., "who works at Altos Labs?" involves person, organization, works-at)
3. Choose strategy:
   • Graph-only — known entity by name/ID, structural question ("who works at X?"), relation traversal
   • Embedding-only — semantic/fuzzy query ("papers about CRISPR delivery"), no specific entity known
   • Hybrid — embedding search to find candidates, then graph traversal for context

Stage B: Execute Plan

Compose and execute queries dynamically using query and search commands:

1. Semantic search (if needed): search --query "..." [--collection <name>]
   • Returns entity IDs + scores from Qdrant
   • Cross-collection search (omit --collection) searches ALL embedded collections
2. Entity lookup: query --typeql 'match $e isa <type>, has id "<id>"; fetch { ... };'
3. Relation traversal: Follow connections from entities

   match (employee: $p, employer: $o) isa works-at; $o has name "Altos Labs";
   fetch { "person": $p.name, "person-id": $p.id };
   

4. Notes/claims about entity: match (note: $n, subject: $e) isa aboutness; $e has id "<id>"; fetch { "content": $n.content };
5. Episode history: match (session: $ep, subject: $e) isa episode-mention; $e has id "<id>"; fetch { "episode": $ep.content };
6. Alias resolution: Check entity-alias for cross-skill duplicates

Stage C: Organize Response with Provenance

Synthesize results into a structured answer:

• Direct answer — facts addressing the question
• Supporting evidence — which notes, artifacts, papers support each fact
• Provenance — how each piece was obtained (skill, timestamp, source)
• Confidence — from memory-claim-notes and evidence-chain attributes
• Cross-domain connections — related entities from other namespaces

Schema Navigation Patterns

Understanding the type hierarchy

describe-schema returns the full entity/relation hierarchy with:

• parent: direct supertype (e.g., person's parent is agent)
• owns: all attributes including inherited
• plays: roles this type can play in relations
• subtypes: types that inherit from this type
• embedding_index: which Qdrant collections have embedded vectors for which types

TypeDB's subtype polymorphism is key: match $p isa person returns ALL person subtypes (operator-user, author, jobhunt-contact, etc.). Use the broadest useful supertype for queries.

Finding connections between entities

1. Check what relations the entity's type plays (from describe-schema)
2. Query each relevant relation:

   match $e isa person, has id "<id>"; (employee: $e, employer: $o) isa works-at;
   fetch { "org-id": $o.id, "org-name": $o.name };
   

3. For notes/analysis: use aboutness relation
4. For episodes: use episode-mention relation
5. For collections: use collection-membership relation

Cross-skill entity visibility

The same real-world entity may exist as different TypeDB types across skills. Use entity-alias to link them, and check aliases when building unified views.

TypeQL Patterns Cheat Sheet

-- Polymorphic match (catches all subtypes)
match $p isa person; fetch { "id": $p.id, "name": $p.name };

-- Relation traversal
match (employee: $p, employer: $o) isa works-at;
fetch { "person": $p.name, "org": $o.name };

-- Multi-hop: person -> org -> collection
match
    (employee: $p, employer: $o) isa works-at;
    (collection: $c, member: $o) isa collection-membership;
fetch { "person": $p.name, "collection": $c.name };

-- Notes about an entity
match
    $e isa identifiable-entity, has id "<id>";
    (note: $n, subject: $e) isa aboutness;
fetch { "note-id": $n.id, "content": $n.content, "confidence": $n.confidence };

-- Text search (substring match)
match $e isa person, has name $n; $n contains "Burns";
fetch { "id": $e.id, "name": $n };

-- Temporal filtering
match $r (employee: $p, employer: $o) isa works-at, has valid-from $vf;
fetch { "person": $p.name, "org": $o.name, "since": $vf };

Semantic Search

The search command embeds query text with Voyage AI and searches Qdrant collections.

Available collections (check describe-schema for current list + point counts):

• alhazen_papers — scientific papers (title + abstract)
• apt-notes — disease mechanism evidence notes
• jobhunt-opportunities — job positions with sensemaking notes
• dismech_benchmark — disease mechanism corpus (25K+ points)

Cross-collection search (omit --collection) searches ALL collections and merges results by score. This is the "search everything" capability.

Results include entity IDs that bridge back to TypeDB for graph follow-up.

Write Operations

Memory Claims

# Consolidate a fact
agentic_memory.py consolidate --content "..." --subject <entity-id> --fact-type knowledge --confidence 0.9

# Recall claims about an entity
agentic_memory.py recall --subject <entity-id>

# Invalidate (soft delete with timestamp)
agentic_memory.py invalidate <claim-id>

Episodes

# Create with skill attribution
agentic_memory.py create-episode --skill <name> --summary "What happened"

# Link entities with operation tracking
agentic_memory.py link-episode --episode <ep-id> --entities <id1>,<id2> \
  --operation-type created --rationale "Why this was done"

# View structured session history
agentic_memory.py show-episode <ep-id>

Entity Aliases

# Link two entities as the same real-world thing
agentic_memory.py merge-entities --canonical <id1> --alias <id2> --description "Same person"

# Remove alias
agentic_memory.py unmerge-entities --canonical <id1> --alias <id2>

# List all aliases (or for a specific entity)
agentic_memory.py list-aliases [--id <entity-id>]

Operator Context

agentic_memory.py create-operator --name "..." --identity "..." --role "..."
agentic_memory.py update-context-domain --person <id> --domain goals --content "..."
agentic_memory.py get-context --person <id>
agentic_memory.py link-project --person <id> --collection <id>
agentic_memory.py link-tool --person <id> --entity <id>
agentic_memory.py link-person --from-person <id> --to-person <id> --context "..."

Command Output Pattern

All commands return JSON to stdout. uv run emits a VIRTUAL_ENV warning to stderr. Always use 2>/dev/null when piping to a JSON parser.

Read USAGE.md for full command reference.