[Pentexa]

Academy Memory v0.7 — Persistent Causal Memory

Summary

Audit v0.6 confirmed that retained memory does not demonstrate genuine
zero-shot knowledge transfer and that cross-task negative evidence
destroys previously verified hypotheses. v0.7:

• Pins the root cause (Mission A): the global, cross-task
  contradiction accumulator. The agent's context signature is the
  episode-start state (identical across tasks), and the consecutive-
  nonevent window used to count contradictions is never reset at task
  boundaries. A hypothesis verified in task A is contradicted in task B
  (different hidden mechanism, same observable context) and permanently
  marked unusable — even when the same rule recurs in task C.
• Introduces a minimal bounded causal memory (Mission B):
  memory_episode_scoped — per-episode support/contradiction records,
  status computed relative to the active epoch only (evidence-based
  promotion/demotion, no global replacement), epoch-boundary reset of
  the nonevent window, and cross-epoch uncertainty-aware retrieval
  (hypotheses verified in ≥2 past epochs, untested in the active epoch).
• Instruments memory-to-action attribution (Mission C): per-run
  counters for hypothesis retrieval, cross-epoch retrieval, memory-
  driven actions, correct/incorrect memory-driven decisions, and a
  memory-attributable improvement metric.
• Controlled evaluation (Mission D): fresh, original (v0.6
  baseline), memory_wiped, atomtx (v0.6 atom-transfer baseline),
  shuffled-memory negative control, and epoch (v0.7). Evaluation
  configuration is that of v0.6 (eval seed 20261108, 16-config classic
  grid, 12-config OOD, 8-config retention substrate; no new eval seeds).
  Registration note: the Quality v0.7 transfer protocol
  (transfer-eval-protocol-v01) was frozen by the audit after these
  runs; this experiment is therefore reported as preliminary
  evidence (protocol level_1_preliminary), not as protocol-
  preregistered or confirmatory.

Measured results (2 seeds × 160 episodes, frozen 16-config grid, pinned hash)

Variant	Frozen	Adapt	OOD frozen	Continual OOD
fresh	0.312	0.695	0.333	0.333
original	0.281	0.594	0.333	0.500
memory_wiped	0.281	0.594	0.333	0.500
atomtx	0.250	0.617	0.333	0.500
shuffled (negative control)	0.281	0.594	0.333	0.500
epoch	0.594	0.703	0.458	0.333

Paired frozen deltas (candidate − baseline):

Comparison	Δ	95% CI
epoch vs fresh	+0.285	[0.063, 0.531]
epoch vs original	+0.315	[0.094, 0.531]
epoch vs shuffled	+0.315	[0.094, 0.531]
epoch vs novocab	+0.134	[−0.047, 0.320]
shuffled vs original	0.000	[0, 0]

Controls: random_walk 0.050, novocab 0.463 (same frozen grid).

Honest statement (preliminary)

v0.7 offers preliminary memory-dependent transfer evidence, not yet
independent confirmatory proof:

• frozen 0.594 vs fresh 0.312 (Δ +0.285, CI [0.063, 0.531]) — 2 train
  seeds (pilot-grade per protocol v01), CI excludes 0 for this primary
  delta;
• shuffled-memory negative control (same training, broken associations)
  loses the gain entirely (0.281; Δ vs shuffled +0.315, CI
  [0.094, 0.531]) — the memory-attributable component;
• against the exploration-only novocab control the delta is positive
  (+0.134) but its CI includes 0 — a planning-only contribution cannot
  be fully excluded at 95%; the claim is therefore preliminary, not
  confirmatory (confirmatory would require ≥5 train seeds and
  independent audit recomputation);
• negative results preserved: all v0.6 baselines ≤ fresh; atomtx and
  shuffled do not transfer.

Memory diagnosis (confirmed)

• Cross-task contradictions previously destroyed verified hypotheses
  (~75/77 contradicted at checkpoint); epoch reduces contradicted to
  ~15 and raises verified to ~10, with 47 hypotheses verified in ≥2
  epochs (transferable).
• Attribution: 23/32 frozen cells show memory-driven actions (3568
  actions). Frozen cells disable learn(), so door-outcome counters are
  0 by construction there; action-side attribution and success splits
  are measurable, and the shuffled ablation is the decisive control
  (see docs/MEMORY_V07.md §4.1 limitation).

Quality v0.7 checks

• Verifier (incl. --transfer): 52 PASS / 0 FAIL / 4 INCONCLUSIVE
  (memory-limit; format-level c_missing_duplicate_records;
  transfer.unseen.separation — resting on declaration; transfer.
  attribution.consistency — no merged frozen_trace_mean block).
• Protocol checks PASS: protocol registered, roles resolve
  (fresh/original/memory_wiped/shuffled/novocab), frozen immutability
  attested + trace artifact, ablation paired, budgets.

Testing

• python -m unittest academy.tests.test_memory_v07 → 12 tests OK
• memory suites incl. v0.6/v0.5 → 55+ OK
• Full Academy suite → 246/246 OK
• Audit suite (incl. Quality v0.7 transfer-protocol tests) → 170/170 OK

Deliverables

• academy/agents/memory.py, metargon_agent.py
• academy/tools/memory_experiment_v07.py, memory_merge_v07.py
• academy/tests/test_memory_v07.py; x/run_chunks_v07.py
• docs/MEMORY_V07.md
• results/academy_v01/memory_v07/ (raw artifacts: summary, runs JSONL,
  fresh control)
• generalization_audit/manifests/memory_v07.manifest.json (Quality v0.5)

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.