Agentic Memory

The agentic-memory skill provides TypeDB-backed ontological memory with schema-driven retrieval. It manages operator profiles, long-term memory claims, session episodes, and semantic search across the knowledge graph.

Core Capabilities

Operator identity — Create and manage operator profiles with 10 personal context domains (identity, role, projects, team, tools, communication style, goals, preferences, expertise, decision log)
Long-term memory — Consolidate facts as typed nbmem-memory-claim-note entities with confidence scores and provenance
Session episodes — Record what happened in a session (skill used, entities touched, summary)
Three-stage retrieval — Plan (understand schema) → Execute (compose TypeQL queries) → Organize (structure with provenance)
Semantic search — Query across Qdrant vector collections (papers, notes, opportunities, benchmarks)
Entity aliasing — Merge/unmerge duplicate entities

Data Model

Entity	Purpose
`nbmem-operator-role`	Role bearing for the operator (links person to system)
`nbmem-memory-claim-note`	A persisted fact with fact-type, confidence, provenance
`nbmem-episode`	Session record (skill, summary, timestamp)
`nbmem-entity-alias`	Merge relation for duplicate entities

Memory Claim Fact Types

knowledge — factual information learned
decision — choices made and rationale
goal — objectives and targets
preference — how the operator likes to work
schema-gap — detected modeling gaps

Two-Tier Architecture

Short-term — MEMORY.md files (Claude Code’s built-in memory, fast, session-scoped)
Long-term — TypeDB nbmem-memory-claim-note entities (persistent, queryable, cross-session)

The consolidate command promotes short-term observations into long-term typed memory.

Key Commands

# Operator management
uv run python agentic_memory.py create-operator --name "..." --email "..."
uv run python agentic_memory.py get-context --operator-id <id>
uv run python agentic_memory.py update-context-domain --operator-id <id> --domain goals --content "..."

# Memory claims
uv run python agentic_memory.py consolidate --content "..." --subject <entity-id> --alh-fact-type knowledge --confidence 0.9
uv run python agentic_memory.py recall --subject <entity-id>
uv run python agentic_memory.py list-claims --fact-type knowledge --limit 20

# Episodes
uv run python agentic_memory.py create-episode --skill tech-recon --summary "..."
uv run python agentic_memory.py list-episodes --limit 10

# Schema introspection
uv run python agentic_memory.py describe-schema
uv run python agentic_memory.py query "match $x isa nbmem-memory-claim-note; fetch { ... };"

# Semantic search
uv run python agentic_memory.py search --query "..." --collection alhazen_papers

# Entity management
uv run python agentic_memory.py merge-entities --canonical <id> --alias <id>

When to Use

At session start: get-context to load operator identity
After significant work: consolidate key findings
At session end: create-episode to record what happened
When searching across domains: search for semantic similarity
When debugging schema: describe-schema for TypeDB introspection