Lesson 3 — Consolidation

Duration: ~45 min · Mode: per-student sandboxes (writes on) or a teacher-driven shared project · Instrument: memory structure, memory consolidate --dry-run, and the Memories tab

Learning objectives

1. Explain systems consolidation: episodic detail becomes general knowledge.
2. Describe how a machine can discover structure in memories it was never told about (clustering on relations, similarity, and co-access).
3. Argue for or against human gates on machine-generated knowledge.

The science in one paragraph

Brains keep two learning systems: a fast one that records episodes and a slow one that gradually extracts what episodes share (McClelland, McNaughton & O'Reilly, 1995 — complementary learning systems). New facts consolidate dramatically faster when they attach to an existing schema (Tse et al., 2007). Memory Layer mirrors this: atomic memories are the fast store; a consolidation pass fuses three usage signals — explicit relations, text similarity, and which memories get retrieved together — into one weighted graph, finds communities in it, and a value gate accepts only clusters that are cohesive and either heavily used (salient) or dense but neglected (cold_dense). Each accepted cluster is synthesized into a higher-level insight memory — always as a human-gated proposal. The consolidation page holds the full mechanism and citations.

Prep (5 min)

For the hands-on variant each student needs a sandbox: ./classroom/seed.sh <name> with writes enabled. For the teacher-driven variant, one shared project on the projector is enough. A terminal with the memory CLI (or docker compose exec memory memory …).

Guided activity

1. The hidden structure (10 min). Run the read-only structure view:

memory structure --project classroom

It prints the clusters the deterministic scan discovers right now, with their value-gate metrics (size, density, co-access, activation mass) — plus the insight tree, which is empty on a fresh corpus. Ask students what the members of a discovered group have in common, and which of the three signals (linked, similar, used-together) most plausibly grouped them. Nice detail: two lessons of class queries have been feeding the co-access signal — the class has been creating today's input all week.

2. The value gate (8 min). Not every cluster deserves an insight — consolidating noise adds noise. On the board, list the gate's checks: big enough, cohesive enough, and worth it — either salient (heavily used) or cold-but-dense (compressible). Map to human sleep: the day's rehearsed and the day's coherent-but-unprocessed both consolidate.

3. Dry run, then the real thing (12 min).

memory consolidate --project classroom --dry-run

shows exactly what would be proposed, no LLM involved. With a configured LLM (teacher machine), run it for real: the accepted cluster becomes a pending proposal — a draft insight naming the unifying concept, the tensions between members, and the open gaps. Show the proposal queue:

memory loops memory-proposals --project classroom --status pending

4. The human gate debate (10 min). The system will never write an insight without approval. Split the class: one side argues the gate is essential (hallucination risk, provenance, accountability), the other that it can't scale (who reviews ten thousand insights?). Land on the design's actual position: every durable state change is a reviewable, revertible proposal — the seeded corpus itself contains an insight memory saying exactly that. Have students find it in the Memories tab.

5. Tiers (5 min). Insights are themselves memories: later passes can consolidate insights into higher-tier insights — a schema tree that grows with the project (the reflection-tree idea from Park et al., 2023). Show memory structure again on a mature project if one is available.

Exit ticket

"Describe one 'insight memory' your brain has consolidated — the general lesson, plus two of the specific episodes it was built from."

References

• McClelland, J. L., McNaughton, B. L. & O'Reilly, R. C. (1995). Why there are complementary learning systems in the hippocampus and neocortex. Psychological Review, 102(3).
• Tse, D. et al. (2007). Schemas and memory consolidation. Science, 316(5821).
• Park, J. S. et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior. arXiv:2304.03442 — reflection trees.
• Raghavan, U. N., Albert, R. & Kumara, S. (2007). Near linear time algorithm to detect community structures. arXiv:0709.2938 — the clustering used here.