Module #6 — DA active forgetting + decay-path correctness, with honest falsification - #69
Merged
Conversation
…undle Two independent dopaminergic forgetting circuits (Davis & Zhong 2017; Sabandal et al. 2021), step 1 of the resume contract — GREEN on all DB-free checks, NOT yet verified against live PostgreSQL, NOT for merge. - core/active_forgetting.py: pure two-circuit decisions (permanent Rac1 pressure >= Tp; transient DAMB overlap >= X AND age <= W). - handlers/consolidation/forgetting.py: chronic noisy-OR 1-prod(1-sim_i) over newer overlapping neighbors, acute interferer, pin/sleep gates; permanent -> mark_memory_stale(True), transient -> heat*(1-acute_overlap) (magnitude rides measured salience, no invented constant). - pg_store.search_newer_neighbors: created_at>target KNN (I/O half; noisy-OR aggregation stays in the handler, SRP). - sleep.run_deep_sleep returns replayed_ids (sleep-protection signal); consolidate runs the forgetting cycle AFTER replay. - ablation: ACTIVE_FORGETTING unit (28 -> 29). - tests_py/handlers/test_forgetting_cycle.py: 17 DB-free tests (noisy-OR maths + fake-store wiring), all pass; active_forgetting benchmark PASSED. docs/handoff/da-active-forgetting/ carries the resume checkpoint + verified paper findings (normally outside the repo) for cross-machine continuation. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…its) + benchmark gates Replace the saturating noisy-OR (46% over-fire on 384-dim embeddings) with two independent dopaminergic circuits faithful to the Drosophila literature: permanent Rac1 trace erosion (chronic interference x stage vulnerability, leaky integrator requiring sustained pressure) and transient DAMB retrieval block (recency AND overlap). Constants traced to benchmark, no invented values: PRESSURE_LEAK_LAMBDA=0.85, PERMANENT_ACCUM_THRESHOLD=1.25662, ACUTE_OVERLAP_THRESHOLD=0.575, ACUTE_RECENCY_WINDOW_HOURS=13.0 (source: benchmarks/active_forgetting/run_benchmark.py grid search). Schema: forgetting_pressure_accum REAL column (PG migration + SQLite parity) + store.update_forgetting_pressure_accum; handler persists the leaky accumulator. Benchmark gates (isolated cortex_bench, paired OFF=--ablate vs ON, both --with-consolidation so the forgetting cycle actually fires) — no regression: - LME-S --limit 50: R@10 94.0%->94.0%, MRR 0.857->0.856 (delta 0.0pp/-0.001) - LoCoMo --limit 10: R@10 94.8%->95.0%, MRR 0.832->0.836 (delta +0.2pp/+0.004) - BEAM 100K: no consolidation path, MRR 0.501 no crash (schema/store change safe) Davis & Zhong 2017 (Neuron 95:490-503), Sabandal et al. 2021. 30 tests GREEN, benchmark PASSED (9 falsifiers). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
… 5 core modules Wave-3 provenance audit of 5 modules outside the original 33. Full suite GREEN (3620 passed). One real functional bug, four sets of false paper attributions. - platt_calibration.py: FUNCTIONAL BUG. _smoothed_target swapped class counts (positive used n_neg, negative used n_pos), inverting Platt 1999 Eq.7 / Lin-Lin-Weng 2007 Eq.2. Fixed + regression test. CONVERGENCE_TOL 1e-6->1e-5 and MAX_ITERATIONS=100 sourced to Lin-Lin-Weng Algorithm 1; MIN_SAMPLES relabeled engineering default. - reconsolidation.py: 4 false attributions removed (no values changed). 0.15/ 0.65 not from Osan-Tort-Amaral 2011; PE 0.5 and 30d/0.15 age window are engineering (Lee 2009, Milekic & Alberini 2002 = direction only). Emotional multiplier no longer attributed to Yonelinas & Ritchey 2015 (forgetting-rate effect, not reconsolidation gain - category error documented). - hierarchical_predictive_coding.py: DEEP FIX. Removed ungrounded fixed prior [0.30,0.35,0.35] + duplicate ACh transform (triple-counted ACh). Total free energy now F = sum(pi_i * F_i) derived from neuromodulated precision (Friston 2005 eq.7). Logistic squash relabeled engineering. No prod impact (WRITE_GATE_HIERARCHICAL defaults False). - neurogenesis.py: docstring rescoped (temporal-context heuristic, NOT DG pattern separation); Aimone 2011 demoted to motivation-only; 5 constants relabeled engineering defaults. No behavior change. - memory_ingest.py: importance_boost=1.5 relabeled engineering default; Adcock 2006 demoted to motivation-only (paper reports no ratio). docs/provenance/paper-implementation-audit.md: Wave-3 changelog entry added. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
…kasten) Faithfulness fixes for citation-level misattributions found in the Wave-4 provenance audit. Behaviour-preserving (labels/docstrings only): - write_gate_calibration: the 50% acceptance target is the maximum of the Bernoulli (binary) entropy H(p), i.e. Shannon (1948) — not Jaynes' maximum-entropy principle (Ch.11). The value 0.5 was already faithful; only the citation was imprecise. - condensers: drop the Tse 2007 "schema-congruent consolidation" citation for timeline slot-extraction — a cross-domain category error (Tse is rat hippocampus systems consolidation, no computational result for text slots). - wiki_classifier: the 200–3000 char "atomic note" band is an unsourced engineering default, NOT Luhmann's Zettelkasten (which prescribes no char range); Alexander P2/P4 relabelled "inspired by" (framing only); the score threshold marked unsourced/calibration-pending. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
project_to_2d claimed "spectral embedding of the SR graph" (Stachenfeld 2017) but ran a purely-attractive force-directed relaxation with no eigendecomposition. Replace with the real SR spectral embedding: the eigenvectors of the symmetric normalised adjacency S = D^(-1/2) W D^(-1/2) (which shares the eigenvectors of the SR matrix M = (I-γT)⁻¹, T = D⁻¹W — Belkin & Niyogi 2003), using the two subdominant eigenvectors as the grid-cell-like (x, y) axes. Isolated nodes embed at the origin. Also remove dead code that carried false-attribution defects: - build_co_access_graph (zero importers) held a γ^(dist-1) off-by-one vs the SR definition and a non-SR symmetric back-link (SR is directional). - _SR_DISCOUNT became unused — the spectral embedding is γ-independent (the eigenvectors of M are a function of T alone for any γ∈(0,1)). Adds test_cognitive_map.py (was untested): SR scoring, BFS navigation, and the spectral-embedding faithfulness property (the Fiedler vector separates weakly-bridged co-access clusters). Not in the retrieval ranking path. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
causal_graph claimed the PC algorithm (Spirtes & Glymour 1991) but built a PMI-threshold skeleton with single-variable conditioning and temporal-only orientation — not PC. Replace with a faithful PC in new core module causal_pc: - G² (likelihood-ratio) conditional-independence test over binary entity-presence data; chi-square survival via the regularised upper incomplete gamma (Numerical Recipes §6.2) — no SciPy dep; validated against textbook critical values and scipy to ~1e-14. - PC skeleton with growing conditioning sets and separating-set recording. - v-structure (unshielded collider) orientation; temporal precedence only orients edges PC leaves undirected (background knowledge, never overriding a collider). PC now determines graph structure; PMI is demoted to an effect-size annotation for ranking only. discover_causal_edges takes a per-memory presence matrix (build_presence); the cls consolidation handler is updated. Removes dead PMI helpers (compute_conditional_independence, compute_co_occurrence_matrix). Tests rewritten with product/grid designs that make conditional independence exact. Not in the retrieval ranking path. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
Zetetic warning cleanup — each is a real unverified behaviour, not noise: - FutureWarning (production): embedding_engine used the deprecated get_sentence_embedding_dimension; prefer get_embedding_dimension with a getattr fallback for sentence-transformers <5 (cleared ~15 occurrences). - un-awaited coroutine: test_pipeline_impact_bump patched asyncio.run but still evaluated the real _pipeline_detect_changes coroutine as its arg; also patch the coroutine factory (was GC-misattributed to other tests). - leaked psycopg connections: tests constructed MemoryStore(...) directly instead of get_shared_store() (the documented rule — closed by conftest's reset_shared_store teardown). Converted spell/codebase-alteration, pg_store_delete_by_tag, seed_project, remember; closed the schema_integrity bootstrap store. - leaked asyncio subprocess transports: the mcp_client deadlock-regression tests spawn real children; close() (sync) sends SIGTERM but cannot await exit, so reap the child (await proc.wait() + transport.close()) within the test loop before it tears down. Remaining 6 are GC-misattributed psycopg ResourceWarnings to trace later. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
A3 made HEAT lazy (effective_heat decays from heat_base_set_at on every read) but left STAGE eager — consolidation_stage advances only in the consolidation handler, never on the recall/read path. Between consolidation passes a trace stayed frozen at its insert stage 'labile' (alpha=2.0, floor=0.0) while its heat decayed under the labile law, collapsing to ~0. Benchmark and import runs trigger no consolidation pass, so the forgetting law could never shape ranking: everything old read as ~0 uniformly. Fix: make STAGE as lazy as HEAT. New PL/pgSQL effective_stage() re-derives the stage at read time from elapsed dwell + stored signal columns, mirroring cascade_advancement.compute_advancement_readiness applied cumulatively (DA gate disabled — encoding-time, unavailable on read; monotonic forward-only; stage_hours consumed as a dwell budget stage-by-stage; reconsolidating/unknown pass through unchanged). effective_heat() now keys BOTH alpha (Kandel 2001) AND the permastore floor (Bahrick 1984, Benna & Fusi 2016) on the derived stage instead of the static column. Also includes the step-1 decay-clock anchor: insert_memory sets heat_base_set_at = COALESCE(provided, created_at, now()) in both pg_store and sqlite_store, so imported/backfilled rows decay from their true age rather than from insert time. Verified on cortex_test (rolled-back probe): fresh/imp0.5 -> 1.00 (early_ltp) 365d/imp0.5/0acc -> 0.05 (late_ltp floor) 365d/imp0.5/5acc -> 0.10 (consolidated floor) 365d/imp0.2/0acc -> ~0 (labile, forgetting preserved) test_pg_effective_stage_parity.py pins the SQL ladder against the Python oracle (grid + monotonicity + off-chain passthrough). I2 allow-list line numbers updated for the heat_base_set_at shift (fix confined to pg_schema, no writer added). Suites: infrastructure+invariants 415 passed. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
Result snapshot from the active_forgetting benchmark run on 2026-06-29, joining its already-tracked siblings in benchmarks/results/active_forgetting/. Predates the lazy-stage read-path fix; recorded for the reproducibility trail. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
effective_heat decayed with exponent α(eff_stage)·hours_elapsed: the final-stage α was applied to ALL elapsed hours. At a stage boundary α drops (e.g. 0.8→0.5), retroactively shrinking the exponent so heat RISES with age — non-monotonic. Latent in every multi-stage trace; the forgetting-curve grid exposed it at the 7h late→consolidated boundary (6h 0.98982 → 8h 0.99151). Replace the α-CASE with a new SQL fn alpha_integral(stage,τ,imp,acc, schema) = ∫₀^τ α(stage(s))ds, walking the SAME chain/gates/dwells as effective_stage. effective_heat now decays over alpha_integral(stage_hours) - alpha_integral(max(0, stage_hours-elapsed)). Since α>0 ⇒ ∫α strictly increasing ⇒ POWER strictly decreasing, heat is monotone non-increasing BY CONSTRUCTION. Single-stage traces reduce exactly to α·hours_elapsed (max|Δ|=0.0 vs the old path, verified). Also fix emergence_metrics.compute_forgetting_curve docstring: it claimed a power-law R=a·t^-b fit, but the code fits ln(heat) on linear age — an EXPONENTIAL. curve_type was already correctly "exponential"; the real power-law fit lives in benchmarks/forgetting_curve/curve_fit.py. Verified: new test_pg_alpha_integral.py (SQL==Python oracle, ∫α non-decreasing, single-stage==α·τ, effective_heat monotone for all 4 profiles); full tests_py/infrastructure/ 380 PASS. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
…ix snapshot New honest falsification instrument for the forgetting law: criteria.py (power-over-exponential AIC, permastore floor, exponent plausibility), curve_fit.py (real power-law fit R=a·t^-b), run_benchmark.py. It does NOT fake a power law — it fits both families and reports which wins. Post-fix snapshot (20260630T102251Z): all 4 stage profiles monotone; C2 permastore PASS (B@365d holds floor, A collapses); C1 still correctly FAILS (ΔAIC favours exponential) — the monotonicity fix removes the stage-boundary bump without manufacturing a power law. Verdict: PARTIALLY FALSIFIED, preserved honestly. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
Post-fix run (20260630T102401Z) confirms no regression from the effective_heat α-integral monotonicity fix: all active-forgetting falsifiers still PASS. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
…(C4) The forgetting-curve harness had a descriptive-only ensemble_diagnostic. Promote it to a falsifiable criterion (C4) that tests the specific Benna&Fusi 2016 prediction: a cascade / continuum of timescales decays as h ∝ 1/√t (power law, exponent ≈ 0.5). C4 is evaluated on the population MIXTURE (mean over the 4 stage profiles), not a single trace — a single stage is a pure exponential by construction, so superposition of separated timescales is the only route to a power law in this architecture. PASSES iff the mixture is fit better by a power law than a single exponential (ΔAIC>2) AND the fitted exponent lies in the √t band [0.4, 0.6]. The double gate matters: in the full run the mixture's fitted b lands at 0.39-0.44 — near 0.5 — yet the power law LOSES to the exponential (winner=exponential, power r²(h)=0.0 because the permastore floor is a plateau no power law can fit). Gating on the exponent alone would have been a false positive. Honest result (20260630T110048Z): C4 PASS=False. The 4-level α-ladder (rates 2.0→0.5, a single decade) is far coarser than Benna&Fusi's many-decade continuum and does NOT reproduce 1/√t. C4 does not gate overall_passed (still C1+C2); it appends an explicit law-family clause to the verdict. C1/C2/C3 unchanged from the prior snapshot (C1 ΔAIC=-124.37) — purely additive, no regression. New unit tests (test_forgetting_curve_criteria.py) lock the falsifier: a planted 1/√t mixture passes, a single exponential fails, and an out-of-band power law (b≈1.5) fails despite winning over exponential. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
…teau The Benna&Fusi √t criterion (C4) fit the mixture over ALL ages via a `1e-3 < h` filter, which included the permastore floor plateau (mean 0.0375 for t≥90d). A power law a·t^-b can represent neither a ceiling at 1.0 nor a floor, so its h-space r² went negative and clamped to 0.0 — a degenerate result that read like a bug and masked the real finding. Fix: restrict C4's fit to the strictly-decaying transient regime via the same transient_points filter C1/C3 already use, with the mixture floor = equal-weight mean of the per-profile permastore floors. The plateau is a separate phenomenon (Bahrick, tested by C2), not part of the decay law. Result is now non-degenerate and cleaner: the transient mixture is an almost-perfect single exponential (exp r²=0.9915) and decisively NOT a power law (power r²=0.230, ΔAIC=-58.5, fit b=0.235 outside the √t band [0.4,0.6]). Verdict unchanged: the 4-level α-ladder does not reproduce the 1/√t continuum. Output now reports n_transient_points + mixture_floor. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
…(Wave 3 sibling) predictive_coding_flat.py had the same zetetic violation Wave 3 fixed in its hierarchical sibling: invented constants under the Friston 2005 label with no provenance disclaimer (the honesty discriminator is the disclaimer, not the citation). The flat path was missed in the first Wave 3 pass; surfaced during the forgetting-law campaign. No values changed (behavior-preserving). Every constant now carries a source/disclaimer: the 4-signal blend weights (0.40/0.25/0.20/0.15), the 24h temporal saturation timescale, the order-of-magnitude length buckets, and the no-data fallback priors (0.5/0.7/0.8) are relabeled engineering defaults; Friston 2005 demoted to concept-only (prediction-error gating). The blend is validated end-to-end — ROC-AUC=0.9998 separating novel from duplicate content (flat mode, benchmarks/gate_precision, 2026-06-11; the hierarchical alternative scored only 0.5514, which is why flat is the default path). Audit doc Wave 3 changelog updated with this entry. Verified: 156 write-gate/predictive-coding/novelty/gate tests PASS. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
…-decayed heat Remove the age_decayed_heat pre-decay override from backfill_memories and import_sessions. Post-A3 (382f394), insert_memory anchors heat_base_set_at to created_at, so effective_heat() already decays the baseline by the memory's real age at read time. Pre-decaying initial_heat on top double- counted the same age — verified on cortex_test as a 1.05x-1.83x effective_heat suppression in the 3-60 day window (masked by the late_ltp floor beyond ~90d). Keep created_at; drop the now-dead age_decayed_heat/compute_age_days helpers and their unit test. The original #14 bimodality motive is subsumed by A3's read-time decay; the homeostatic cohort correction (#14 Fix 2) is untouched. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
…ine retrieval The benchmark table's retrieval scores are measured with consolidation enabled. The runners default to consolidation OFF, where every loaded memory keeps its ingest-time heat timestamps (hours_elapsed≈0), so the stage never advances past labile, effective_heat sits at the labile floor (≈1e-4), and the read-path heat gate (min_heat=0.01) prefilters every candidate — retrieval collapses to ≈0%. This is a harness artifact of the OFF default that reproduces across branches, not a property of the stored memories or of this branch. Subset reconfirmation on cortex_bench (2026-06-30, --with-consolidation vs OFF): LongMemEval-s n=50 R@10 94.0% / MRR 0.854 vs 0%; LoCoMo n=10 (1982 Qs) R@10 93.8% / MRR 0.821 vs 9.3%. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
The branch's new/changed Python files were committed without running `ruff format`; CI's Lint job (`ruff format --check .` + `ruff check .`) failed. Apply ruff formatting to the 17 affected files (formatting-only, AST-preserving) and remove the unused `forgetting_pressure` import in the active_forgetting benchmark (F401). `ruff format --check` and `ruff check` both pass locally. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au
Merged
cdeust
added a commit
that referenced
this pull request
Jun 30, 2026
…forgetting (#73) Bumps version across pyproject/server/manifest/marketplace/plugin manifests and records the 3.25.0 CHANGELOG section covering #67–#72: - feat: active forgetting module (#69), safe headless drain (#70), full zetetic agent roster for wiki authoring (#72) - fix: Windows portability (#68), subscription-default billing (#71), variadic --add-dir prompt-swallow root cause (#72) - deps: pydantic-settings 2.14.0→2.14.2 security patch (#67) Claude-Session: https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
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.
Module #6 — DA active forgetting + decay-path correctness, with honest falsification
This branch lands the active-forgetting mechanism (two Drosophila dopaminergic
circuits) and a campaign of correctness and provenance fixes around the
memory decay path. It is shipped with a falsification harness that we let
fail where the model genuinely diverges from biology rather than tuning the
test to pass.
What landed
Feature — active forgetting (
a5bef4e8,c264af95)core/active_forgetting.py+handlers/consolidation/forgetting.py: twoindependent dopaminergic forgetting circuits — permanent Rac1 trace erosion
(chronic interference × stage vulnerability) and transient DAMB retrieval
block (Davis & Zhong 2017; Sabandal et al. 2021).
benchmarks/active_forgetting/(9/9 pass) andLME on/off gate snapshots.
Decay-path correctness
42e6ff19— backfill/import use the single A3 decay path: removed theage_decayed_heatoverride that double-decayed heat (created_at → A3 decayonly). Deleted dead
backfill_helpers.pyandtest_backfill_heat.py.acfc948c—effective_heatmade monotone via a piecewise α-integral.382f3948— consolidation stage derived lazily on the read path.Provenance corrections (zetetic source discipline)
f8b5472d,ee201ff1— corrected false paper attributions(Shannon / Tse / Zettelkasten / Platt Eq.7) across 5 core modules.
eaca9636— constant disclaimers onpredictive_coding_flat(engineeringdefaults, not paper-derived).
ca545438— real PC algorithm (G² CI test + v-structures) replacing theplaceholder.
9c7cc373— faithful SR spectral embedding for the cognitive map; dead codedropped.
Transparency — falsification preserved, not papered over
The forgetting-curve harness (
benchmarks/forgetting_curve/, criteria C1–C4)is falsifiable by construction and we report its verdict honestly:
Ebbinghaus with a floor, not a power law.
transient regime, not a permastore √t plateau (
12f6637b,04357868).exponential+floor model, and the harness says so out loud rather than being
retuned to confirm a law the system does not implement.
This matches the thermodynamic paper, whose decay axis is exponential
Ebbinghaus (
R(t)=e^{-t/S}); the paper's "power law" refers to thecross-sectional steady-state item distribution, not a retention curve.
Independent paper audits
Both arXiv companion papers were audited against the implementation on all 7
campaign axes (exp-not-power-law; Adcock 1.5×; reconsolidation 0.15/0.65;
Yonelinas forgetting-rate-not-gain; no fixed prior; neurogenesis≠DG;
engineering-default constants):
docs/arxiv-thermodynamic/main.tex— FAITHFUL, zero HIGH/MEDIUMcontradictions.
docs/arxiv-context-assembly/main.tex— FAITHFUL, zero contradictions.Benchmarks — no regression on the heat gate
The retrieval heat gate is byte-identical between
mainand this branch(the
effective_heatprefilter appears 5× inpg_schema.py, unchanged on bothsides; the changed code is the backfill/import path, off the benchmark
ingestion route). So retrieval scores are governed by the same code on both
branches.
Consolidation is required for headline retrieval. The benchmark runners
default to consolidation OFF; in that mode every loaded memory keeps
heat_base_set_at/stage_entered_atat ingest-time (hours_elapsed ≈ 0), sothe stage never advances past labile,
effective_heatsits at the labilefloor (
≈1e-4), and the read-path gate (min_heat = 0.01) prefilters everycandidate → retrieval collapses to ≈0%. This is a harness artifact of the
OFF default, not a property of the branch: it reproduces identically on
main. Reported headline numbers (CLAUDE.md) are measured--with-consolidation.Reconfirmation on an isolated
cortex_benchDB,--with-consolidation(50-Q /10-Q subsets — smaller than the full-suite headline, so not directly comparable
to the 500-Q/full LoCoMo numbers):
--with-consolidationDecay-law benchmarks re-ran with identical verdicts to committed snapshots
(forgetting_curve C1 FALSE / C2 TRUE / C3 TRUE / C4 FALSE; active_forgetting
9/9). 970 tests pass.
🤖 Generated with Claude Code
https://claude.ai/code/session_019o58McF4LRfvGNNXaqG2Au