atmem/laya-memory-decision-selective-v3-research

Laya Memory Decision — Selective V3 Research Candidate

This is a portable, multilingual Laya checkpoint for one bounded memory question: choose STORE_NEW or DO_NOT_STORE from an observation plus authorized same-scope retrieved memory state. It is independent of AtMem: any host can construct the versioned JSON request, keep the opaque option binding, validate the probability vector, and retain final mutation authority.

Outcome: rejected for automatic activation. The checkpoint is public for reproduction and inspection, not as a recommended production model. It predicts DO_NOT_STORE too often and did not improve the matched operational result over deterministic AtMem.

Contract

Input:

{
  "format": "laya-evaluation-request-v2",
  "schema_version": "2.0.0",
  "request_id": "opaque-request",
  "scope_id": "opaque-scope",
  "canonical_generation": 184,
  "observation": {
    "id": "opaque-observation",
    "content": "Actually, call me Sara from now on.",
    "speaker_role": "user",
    "source_digest": "sha256:..."
  },
  "retrieved_memories": [],
  "question": {
    "question_id": "formation.joint_transition.v3",
    "type": "choice",
    "instructions": "Select the one memory-state transition supported by the authorized observation and retrieved active memories.",
    "options": [
      {"id": "option_000", "description": "Create a distinct durable memory from the observation."},
      {"id": "option_001", "description": "Do not create persistent memory from the observation."}
    ]
  }
}

Output:

{
  "format": "laya-choice-response-v2",
  "schema_version": "2.0.0",
  "request_id": "opaque-request",
  "question_id": "formation.joint_transition.v3",
  "type": "choice",
  "choice": "option_001",
  "probabilities": {"option_000": 0.04, "option_001": 0.96},
  "confidence": 0.96
}

Option meanings must be bound by the host request. The model does not write, update, supersede, forget, or delete memory. It does not generate canonical memory text.

Training

• Base: convaiinnovations/laya-multilingual at 1720e3e3357cfe1e281542e223f8273b0890ca34.
• Dataset: atmem/stateful-memory-transitions-selective-v3-research at f46cdcd23c189cacdaa68abd2ec780752227e9f4.
• Objective: soft cross-entropy, one epoch, seed 410239.
• Device: Apple MPS research run; MPS scatter-reduce determinism is unavailable and recorded warn-only.
• Conversation content: authentic source conversations only; no generated conversation, LoCoMo, translation, or sealed-test label was used.
• Exact-packing policy: 218 train, 49 validation, and 58 calibration examples exceeded the 1,024-token serving budget and were excluded before optimization or scoring.
• Retained train coverage: 1,647 DO_NOT_STORE and 50 STORE_NEW distinct trajectories. The positive class is below the 100-trajectory activation minimum.
• Weights SHA-256: 93f1d0c623b956fbe9c92a25b57b4298a08ff8bf5e0b7f182635361d2c304104.
• Training time: 1,136.37 seconds.
• No calibration bundle was issued because semantic coverage and overflow gates failed.

Results

Authentic validation subset that fits the serving budget

Measure	Original Laya	Fine-tuned candidate
Evaluated transitions	423	423
Accuracy	72.34%	96.22%
Macro-F1	0.443	0.627
Brier score	0.3978	0.0727
ECE	0.0830	0.0361
Predicted STORE_NEW	105	4

Only about 3 of 18 positive validation examples were recovered. The apparent aggregate gain is mostly improved DO_NOT_STORE precision, not strong durable memory recall.

Matched frozen 100-item operational replay

Measure	Deterministic	Original Laya	Fine-tuned candidate
Exact policy agreement / 100	30	30	20
STORE correct / 25	25	0	0
DO_NOT_STORE correct / 20	5	0	20
UPDATE correct / 15	0	0	0
DUPLICATE correct / 10	0	0	0
REVIEW correct / 30	0	30	0
Median end-to-end latency	3.15 ms	45.35 ms	45.26 ms
p95 end-to-end latency	4.73 ms	52.55 ms	52.84 ms

The fine-tuned candidate returned DO_NOT_STORE for all 100 operational cases. The operational corpus is a maintainer-created policy challenge, not an independent scientific benchmark; its value here is the matched product regression test.

Intended use

Use this checkpoint to reproduce the negative result, test the strict V3 wire contract, or study selective memory-decision training. Do not enable it as an automatic memory authority. A future candidate needs materially more authentic STORE_NEW coverage, relational transition coverage, tokenizer-budget-safe state packing, per-action calibration, and a matched gain over deterministic formation before activation.