smem: add --sweep-dominated for destructive structural compaction (experimental, changes retrieval behavior)#580
Closed
kimjune01 wants to merge 12 commits intoSoarGroup:developmentfrom
Closed
Conversation
Periodically scan episodic memory for stable WME structures and write
them to semantic memory as new LTIs. This implements the compose+test
framework (Casteigts et al., 2019) for automatic episodic-to-semantic
knowledge transfer — the operation Soar's long-term declarative stores
have been missing.
Algorithm:
compose — union of constant WMEs currently active in epmem
test — continuous presence >= consolidate-threshold episodes
write — create smem LTI with qualifying augmentations via CLI_add
New parameters (all under epmem):
consolidate on/off (default off)
consolidate-interval integer (default 100) — episodes between runs
consolidate-threshold integer (default 10) — min episode persistence
Deduplication via epmem_consolidated tracking table prevents repeated
writes across consolidation runs. Table is dropped on reinit alongside
other epmem graph tables.
Off by default — zero behavior change until explicitly enabled.
Limitations (deferred to follow-up):
- Only consolidates constant-valued WMEs, not identifier edges
- No back-invalidation across the WM/smem tier boundary
- last-consolidation stat does not persist across agent reinit
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.
This patch addresses the first half: automatic semantic learning from
episodic experience. With semantic entries derived from episodes,
episodic eviction becomes safe (merged episodes leave no reconstruction
debt), and R4's forgettable WME scope expands automatically.
Reference:
Casteigts et al. (2019), "Computing Parameters of Sequence-Based
Dynamic Graphs," Theory of Computing Systems.
Derbinsky & Laird (2013), "Effective and efficient forgetting of
learned knowledge in Soar's working and procedural memories,"
Cognitive Systems Research.
https://june.kim/prescription-soar — full prescription
After consolidation writes stable WMEs to smem, old episodes become redundant. Delete point entries and episode rows older than consolidate-evict-age episodes. This is safe: the consolidated knowledge is in smem, so there is no reconstruction debt. New parameter: consolidate-evict-age integer (default 0 = off) — min age before an episode is eligible for eviction Range and _now interval entries are preserved (they span multiple episodes). Only point entries and episode rows are removed. Reference: Derbinsky & Laird (2013), §5 — "forgotten working-memory knowledge may be recovered via deliberate reconstruction from semantic memory." Consolidation creates the semantic entries; eviction removes the source episodes that are no longer needed for reconstruction.
- Delete _range entries whose intervals end before the eviction cutoff (previously only _point entries were evicted, leaving dead weight) - Wrap all eviction DELETEs in BEGIN/COMMIT when lazy_commit is off for atomicity (when lazy_commit is on, already inside a transaction) Retrieval of evicted episodes is already safe: epmem_install_memory checks valid_episode and returns ^retrieved no-memory.
Implements Kilpeläinen-Mannila tree inclusion to detect which LTI entries are structurally dominated by others. Detection only — no eviction. Works at the raw hash level via web_expand to avoid Symbol allocation overhead.
…rror routing Codex review found four issues: 1. Greedy child matching could produce false negatives when first-fit blocks later required matches. Replaced with backtracking injective matcher. 2. Cycle detection conflated "currently exploring" with "proven to include". Split into separate active_pairs (recursion stack) and memo (proven results). Cycles now conservatively return false. 3. CLI_redundancy_check returned void, DoSMem always returned true. Changed to bool with SetError routing on failure. 4. help smem did not list --redundancy-check. Added.
Round 2 codex review found two remaining issues: 1. Per-attribute a_used vectors allowed two distinct B nodes to map to the same A node via different attributes. Fixed by threading a global b_to_a map (B node → A node assignment) through all recursion. Backtracking undoes assignments on failure. 2. Active-pair cycle check returned false (pessimistic), which missed valid cyclic equivalences like @1 ^next @1 vs @2 ^next @2. Changed to coinductive (optimistic): revisiting an active pair returns true. If the assumption is wrong, non-cyclic proof obligations will fail.
Round 3 codex review found two issues: 1. Failed speculative branches leaked descendant b_to_a bindings. Fixed by snapshotting b_to_a before each branch and restoring on failure. Also pre-bind b_child before recursion so descendants see the intended assignment. 2. Memoization keyed by (lti_a, lti_b) was unsound because results depend on the current b_to_a context. Dropped memo entirely — smem entries are shallow (depth 1-2) so re-evaluation is cheap.
Round 4 codex review found two issues: 1. Root of B was not pinned to root of A in the global assignment. Counterexample: @b ^next @b vs @A ^next @A1, @A1 ^next @A1 — B's root could map to A1 instead of A. Fixed by seeding b_to_a[lti_b] = lti_a before recursion. 2. smem --redundancy-check was missing from the runtime help screen in smem_settings.cpp. Added.
…ated LTIs Mark phase reuses tree inclusion detection from SoarGroup#579. Sweep phase: 1. R4 safety check: skip LTIs currently referenced in working memory 2. Dependency-safe deletion: disconnect_ltm, then delete from all smem tables (augmentations, activation history, aliases, lti) 3. Per-invocation budget via optional numeric argument Includes functional test proving full eviction: add 3 LTIs where @1 is structurally dominated by @2, sweep, verify @1 is gone and @2/@3 survive. Post-sweep redundancy check confirms no remaining dominated entries.
Replaces raw SQL deletion in sweep with a proper kernel-side routine that composes existing bookkeeping paths: 1. disconnect_ltm for outgoing edges (existing) 2. Inbound edge update: for each parent pointing to this LTI, decrement child counts, LTI-child counts, attribute frequencies 3. Invalidate spreading activation via invalidate_from_lti 4. Clean all 8+ auxiliary tables: prohibited, trajectories, likelihoods, trajectory_num, current/uncommitted/committed spread, activation history, aliases, fake activations 5. Delete from smem_lti and update node count Addresses all three correctness blockers from codex round 1 review of the sweep PR.
Round 2 codex review found: remaining_attr_q excluded constant rows because NULL <> ? is NULL in SQLite (not true). A parent with ^name @dead AND ^name alice would incorrectly decrement attribute frequency after deleting @dead. Fixed with (value_lti_id IS NULL OR value_lti_id<>?). Added testSweepDominatedWithInboundRefs: parent @1 has ^name alice (constant) and ^friend @2 (LTI child). @2 is dominated by @3 and swept. Verifies @1 retains ^name after the LTI child is removed.
… test Round 3 codex review found: (1) surviving parents need trajectory invalidation and edge weight renormalization after child removal, (2) NULL-safe attribute frequency check already fixed. Guard spreading invalidation on spreading-enabled check to avoid hangs when spreading tables aren't initialized. Added testSweepDominatedWithInboundRefs: parent with LTI child where child is dominated and swept. Verifies parent survives with constant attributes intact, no redundancy remains after sweep.
kimjune01
changed the title
kimjune01
changed the title
Author
|
Retiring this PR. The assumptions it was built on were wrong:
See the re-intake at june.kim for the corrected SOAP analysis. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this PR does
Adds
smem --sweep-dominated [<budget>]: a destructive CLI command that deletes LTIs identified by the same structural inclusion predicate used bysmem --redundancy-check, up to an optional budget. Also adds a kernel routinedelete_ltm()used as the deletion path.Off by default. Must be invoked explicitly. Deletion is irreversible within the store; this command is intended for manual maintenance workflows, not routine agent execution.
Important: this command changes smem behavior, not just storage
Structural containment under the inclusion predicate does not imply retrieval equivalence. Deleting a "contained" LTI can change retrieval results qualitatively, not just shift ranking:
Strict behavior preservation would at least require merging activation history and spreading contributions from the removed LTI into its container, and likely additional semantic reconciliation beyond that. This PR does not attempt any of it. Retrieval results and activation dynamics may change in ways this command does not predict.
Algorithm
ids in working memory viasmem_in_wmem. This filter is best-effort only. It does not track LTIs referenced as WMEvalues, so a leaf LTI present in working memory only as a value may not be protected.delete_ltm()for each contained entry in reverse ID order, up to the optional budget.delete_ltm(pLTI_ID): centralized deletion bookkeepingConsolidates the bookkeeping currently needed by this command into one routine. The steps below are what is currently performed; this list is not a claim of completeness against the entire smem schema.
disconnect_ltm.smem_prohibited,smem_trajectories,smem_likelihoods,smem_trajectory_num, the spread tables used by the current spreading implementation, activation history, aliases, and fake activations.smem_ltiand updatesltinode count statistics.If other auxiliary tables or invariants touch the deleted ID and are not listed above, they are not being maintained by this routine. A maintainer review of the full schema is a prerequisite for using this command in production.
Known limitations
smem_in_wmemtracks LTIs as WMEids only, not asvalues. A leaf LTI present in working memory only as a WME value can slip through the filter and be deleted while still referenced. This is best-effort, not a safety guarantee.Test plan
testSweepDominated: flat entries, verify deletion and survivaltestSweepDominatedWithInboundRefs: parent with LTI child, child contained and deleted, parent retains constant attributes