Skip to content

Background Learning Engine (CONCEPT:KG-2.13)

Overview

The Background Learning Engine runs an asynchronous, concurrency-bounded learner that turns conversation transcripts into targeted ADD / UPDATE / DELETE memory edits — not raw dumps — and writes them as bi-temporal graph mutations (KG-2.11). Assimilated from Quarq Agent's async learner (agent-oss/agent.py), with a memory-os-inspired typed, outcome-grounded extraction enhancement. Extends KG-2.1 (+AHE-3 self-improvement).

How it works

  • Targeted edits. extract_edits (ORCH-1.27 learner role) emits a MemoryEdit list: ADD new facts, UPDATE outdated ones, DELETE contradictions.
  • Typed + grounded (enhancement). Each edit carries entry_type (decision/resolution/note/fact), a training_value, an outcome_gate, and evidence_ids. A decision/resolution that claims grounding but cites no evidence is dropped (not stored as an unverified fact); persisted decisions get GROUNDED_BY edges to their evidence and type:/train: tags.
  • Bi-temporal mutations. apply_edits ADDs with full temporal stamps; UPDATE re-stamps and supersedes; DELETE is soft (status=REMOVED + valid_to), preserving history.
  • Async controls. Semaphore(4), bounded exponential backoff (with_backoff, 2→60s, capped so CI never hangs), schedule + await_pending sync barrier. resolve_relative_dates converts "yesterday"/"N weeks ago" to absolute dates at learn time.

Key files / API

Piece Location
Engine + edits knowledge_graph/memory/learning_engine.py (MemoryEdit, BackgroundLearner, extract_edits, run_learner, resolve_relative_dates, parse_memory_edits, with_backoff)
Entry point knowledge_graph/memory/cli.py (agent-utilities-memory learn)

Wiring (≤3 hops)

agent-utilities-memory learn → run_learner → engine.add/update/delete_memory_node (3 hops, verified).

Research provenance

Quarq async learner — agent-oss/agent.py:99-160, 2951-3007, 3303/3646; typed-extraction enhancement — memory-os icarus/hooks.py (ClaudioDrews/memory-os@a4ca094).