MMLA Memory

中文 README

MMLA studies how a bounded memory system can let earlier reasoning influence later reasoning without blurring causal order, authority, or evidence. The technology tree separates a validated component record from registered negative results and from five formal successor branches: reasoning-time training, atomic memory rows, predictive admission, dual policy/memory state, and completed-segment consolidation.

• Public paper: arXiv:2606.28876
• Project repository: MMLA-org/mmla-memory
• Authors: Junyi Zou and Avrova Donz (equal contribution)

Documents

• Complete Technical Report — the additive V3-R24 candidate preserves the complete R22 evidence record and integrates all five formal branches, their proofs, counterexamples, costs, and stop rules.
• Reasoning-Time Training — defines a typed within-problem process and the evidence needed to distinguish policy-state learning from search, context, memory, or workspace effects.
• Atomic Memory Rows — develops a bounded typed row, trusted assembly, exact commit/NULL behavior, lifecycle semantics, recovery, and explicit resource accounting.
• Predictive Memory Admission — formalizes event-indexed grouped futures, expected-risk comparators, exact NULL, uncertainty-aware admission, identification limits, and falsification gates.
• MMLA-RTT Dual State — keeps policy state and authoritative memory distinct in type, privilege, lifetime, reset, rollback, and ledger while specifying conditional composition and identification.
• Completed-Segment Consolidation — gives a causal post-closure contract with five clocks, immutable emitted history, unique event ownership, teacher separation, exposure rules, and bounded work.

Evidence status

Validated component evidence

The public V3/R22 record contains the admitted component evidence: controlled lifecycle behavior, sparse calibrated retrieval/fallback behavior, and typed transport and resident-state components under their stated scopes.

Registered negative results

The same record retains the closed scientific gates: the PM-I2 semantic interface did not qualify (0/9), PM-I3 found fitted-support failure for the learned latent-row route, and learned predictive overwrite remains unvalidated.

Theory and protocol successors

The five focus papers and V3-R24 are local candidates pending independent Reviewer adjudication. They add conditional mathematics, systems contracts, counterexamples, and unrun protocols; they contain no new experimental result and do not establish an oracle gap, learned admission policy, strict-RTT or CSBC success, safety, or end-to-end efficiency superiority.

Citation

Please cite the public arXiv paper as the primary public record:

@article{zou2026mmla,
  title         = {MMLA: How Memory Lets the Past Shape the Future},
  author        = {Zou, Junyi and Donz, Avrova},
  year          = {2026},
  eprint        = {2606.28876},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI}
}

The following local candidate records are separate from the public arXiv publication:

@techreport{zou2026mmlar24,
  title  = {MMLA: How Memory Lets the Past Shape the Future},
  author = {Zou, Junyi and Donz, Avrova},
  year   = {2026},
  note   = {V3-R24 complete technical report candidate; pending independent review}
}

@techreport{zou2026rtt,
  title  = {Reasoning-Time Training: Learning Before a Single Problem Ends},
  author = {Zou, Junyi and Donz, Avrova},
  year   = {2026},
  note   = {MMLA focus-paper candidate; pending independent review}
}

@techreport{zou2026amr,
  title  = {Atomic Memory Rows: A Bounded, Verifiable Substrate for Editable Reasoning},
  author = {Zou, Junyi and Donz, Avrova},
  year   = {2026},
  note   = {MMLA focus-paper candidate; pending independent review}
}

@techreport{zou2026pma,
  title  = {Learning What to Remember: Predictive Admission for Bounded Reasoning-Time Memory},
  author = {Zou, Junyi and Donz, Avrova},
  year   = {2026},
  note   = {MMLA focus-paper candidate; pending independent review}
}

@techreport{zou2026dualstate,
  title  = {MMLA-RTT: Dual-State Learning at Reasoning Time},
  author = {Zou, Junyi and Donz, Avrova},
  year   = {2026},
  note   = {MMLA focus-paper candidate; pending independent review}
}

@techreport{zou2026csbc,
  title  = {Causal Generation, Retrospective Consolidation: Completed-Segment Bidirectional Memory Without Temporal Leakage},
  author = {Zou, Junyi and Donz, Avrova},
  year   = {2026},
  note   = {MMLA focus-paper candidate; pending independent review}
}