Episodic-to-semantic memory consolidation (experimental)

[kimjune01]

Summary

Experimental / demonstration only. This PR adds episodic-to-semantic consolidation and episodic eviction — steps 1 and 2 of the Prescription: Soar. Together they close the loop that Derbinsky & Laird (2013) left open: semantic memory learns from experience, and episodes that have been consolidated become safely evictable.

Consolidation (step 1)

Periodically scans episodic memory for stable WME structures and writes them to semantic memory as new LTIs via CLI_add. Implements the compose+test framework from Casteigts et al. (2019).

Eviction (step 2)

After consolidation, deletes old episode point and range entries and episode rows past the eviction age. Safe because the consolidated knowledge is in smem — no reconstruction debt. Retrieval of evicted episodes returns ^retrieved no-memory via the existing valid_episode check.

Off by default — zero behavior change until explicitly enabled.

Parameters

Parameter	Type	Default	Description
consolidate	on/off	off	Enable periodic consolidation
consolidate-interval	integer	100	Episodes between consolidation runs
consolidate-threshold	integer	10	Min continuous presence to consolidate
consolidate-evict-age	integer	0 (off)	Min episode age before eviction eligible

Algorithm

Step	Operation
Compose	Union of constant WMEs active in epmem's _now table
Test	Continuous presence ≥ consolidate-threshold episodes
Write	Create smem LTI with qualifying augmentations
Dedup	epmem_consolidated tracking table prevents repeated writes
Evict	Delete point, range, and episode rows older than evict-age
Trigger	Every consolidate-interval episodes

The compose+test scan is O(n) in active WMEs per consolidation run. Casteigts et al. prove this is optimal for any parameter computable through their framework.

Motivation

Derbinsky & Laird (2013) proved forgetting is essential to Soar's scaling but only built it for working and procedural memory. Episodic and semantic memory have no eviction and no capacity bound. The root cause (Diagnosis: Soar): you can only safely forget what you can reconstruct (R4), and semantic memory — the reconstruction target — doesn't learn automatically.

This patch addresses both halves:

1. Consolidation — semantic memory learns from episodic experience
2. Eviction — episodes whose regularities have been consolidated are redundant and can be removed

With these in place:

• R4's forgettable WME scope expands (newly perceived categories get merged into smem)
• Episodic store stays bounded instead of growing at 72,000 episodes/hour
• The 50ms decision-cycle threshold has more headroom

Changes

File	Change
episodic_memory.h	4 params, 1 stat, 7 prepared statements, function declaration
episodic_memory.cpp	Param/stat init, SQL statements, schema table + drop, epmem_consolidate() with eviction (~200 lines), hook in epmem_go()
EpMemFunctionalTests.hpp	3 new test declarations
EpMemFunctionalTests.cpp	3 test implementations
2 new .soar test agents	Consolidation and eviction test agents

Tests

46/46 epmem tests pass (43 existing + 3 new), zero regressions:

• testConsolidation — stable WMEs written to smem after consolidation fires
• testConsolidationOff — smem stays empty when feature is disabled
• testConsolidationEviction — old episodes evicted, recent episodes preserved, smem entries intact

Known limitations (deferred)

• Constant WMEs only — identifier (graph-structured) WMEs are not consolidated; consolidated LTIs are flat
• No back-invalidation — cross-tier cache coherence when smem entries become stale (prescription step 3)
• _now intervals preserved — active WME intervals are never split; only completed ranges and points are evicted

Reference

• Casteigts et al. (2019), "Computing Parameters of Sequence-Based Dynamic Graphs"
• Derbinsky & Laird (2013), "Effective and efficient forgetting of learned knowledge"
• Laird (2022), "Introduction to the Soar Cognitive Architecture"
• Diagnosis: Soar — gap analysis
• Prescription: Soar — algorithmic detail

---

See also: #577 — RL convergence gate for chunking (prescription step 4)

[scijones]

Wow, this and the other pull request. I need to review these, but thanks! These look exciting.