Skip to content

Knowledge Layer

The Knowledge layer provides persistent storage across three tiers, each serving a different cognitive function.

Three Knowledge Tiers

Tier System Purpose
Tier 1: Experience Episodic Memory Raw episodes from agent interactions, anchored with provenance metadata
Tier 2: Records Ship's Records Consolidated knowledge — agent notebooks, duty logs, Captain's Log
Tier 3: Operational State KnowledgeStore System state persistence — trust snapshots, routing weights, agent source code

KnowledgeStore is operational state, not shared knowledge

Despite its name, KnowledgeStore stores system operational state (trust snapshots, routing weights, agent source code), not shared knowledge. Ship's Records is the actual shared knowledge fabric where agents write and read knowledge.

Ship's Records

Git-backed instance knowledge store (AD-434). Three components:

  • Agent Notebooks — each crew agent maintains a personal notebook for observations, analysis, and insights. Read-before-write deduplication prevents repetitive entries. Dream consolidation merges and refines entries.
  • Duty Logs — operational records of agent activities and system events.
  • Captain's Log — human-authored entries for strategic decisions and directives.

Notebook Quality Pipeline (AD-550-555)

Six-stage pipeline ensuring notebook quality:

  1. Dedup / Read-Before-Write (AD-550) — agents check existing entries before writing
  2. Dream Consolidation (AD-551) — Step 7g merges scattered entries, detects cross-agent convergence
  3. Self-Repetition Detection (AD-552) — extends peer repetition detection to knowledge layer
  4. Quantitative Baseline (AD-553) — automatic quality metrics capture
  5. Convergence Detection (AD-554) — detects when agents from different departments independently reach the same conclusion
  6. Quality Metrics Dashboard (AD-555) — tracks lexical diversity, redundancy rate, entry length trends

KnowledgeStore (Operational State)

Git-backed persistence for system operational state:

  • Versioned: Every change is a commit — full history with rollback
  • Per-artifact rollback: Revert a single artifact without affecting others
  • Warm boot: On startup, the system loads its last known state

Semantic Knowledge Layer

ChromaDB collections provide vector-similarity search across different knowledge types:

Collection Purpose
Agent designs Self-designed agent blueprints
Skills Learned skill templates
Episodes Episodic memory entries
Context Working memory snapshots
Artifacts General knowledge artifacts

Source Files

File Purpose
knowledge/store.py Git-backed operational state persistence
knowledge/semantic.py SemanticKnowledgeLayer (ChromaDB collections)
ships_records/notebooks.py Agent notebook management
ships_records/duty_log.py Duty log entries
ships_records/captains_log.py Captain's Log