AuthorsWalid Saidi
TL;DR
MutMem V2 uses a cryptographically authorized portable recall and mutation protocol to give independently verifiable evidence across 72/72 terminal verdicts and reasons.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
MutMem V2 states that a self-consistent implementation is not yet a portable protocol; object boundaries, byte encodings, trust roots, failure reasons, and terminal states must be explicit.
Without portable verification, reviewers must trust the production database, runtime, or prose description, making authorization, traceability, and mutation integrity hard to independently check.
HOW IT WORKS
MutMem V2 combines a versioned recall-disclosure envelope, cross-object predicates, a portable mutation profile, and independent Node and Python verifiers into a single protocol.
You can think of MutMem V2 like a cryptographic card catalog for agent memory: every recall and mutation is a signed, length-framed card in a Merkle-indexed drawer.
This lets MutMem V2 prove integrity, authorization, and ordered disclosure for persistent memory in ways a plain context window or opaque database cannot.
DIAGRAM
This diagram shows how MutMem V2 processes a recall bundle into deterministic terminal evidence using the portable recall-disclosure protocol.
DIAGRAM
This diagram shows how MutMem V2 evaluates its portable protocol using structural vectors, cryptographic vectors, and a production-derived conformance corpus.
PROCESS
Canonical RECALL path
MutMem V2 starts with the canonical RECALL path, binding method POST and path aimos recall plus canonical request body and nonce.
Recall disclosure envelope
MutMem V2 constructs the recall disclosure envelope with 13 singleton objects and 5r per result, enforcing deterministic membership and order.
Recall predicate graph
MutMem V2 applies the recall predicate graph to check authority, admission, decision composition, per result evidence, and terminal receipt consistency.
Terminal verdict and receipt
MutMem V2 produces a terminal verdict and signed terminal receipt, binding the Merkle root, result count, actor, request, and authority.
KEY CONTRIBUTIONS
Versioned recall disclosure envelope
MutMem V2 defines a versioned recall disclosure envelope with complete membership 13 plus 5r objects, deterministic ordering, and domain separated SHA 256 commitments.
Cross object predicates for authority and evidence
MutMem V2 introduces cross object predicates covering external trust, identity epochs, revocation, effective authorization, per result provenance, occurrence evidence, and signed terminal receipts.
Portable mutation profile with three terminals
MutMem V2 defines a portable mutation profile preserving the native outcome schema and distinguishing authorized transition, signed noop, and occurrence observation terminals.
RESULTS
Portable protocol components
18 schemas; 39 recall vectors; 15 mutation vectors; 37 recall failure codes
Defines full terminal vocabulary and predicates for MutMem V2
Verifier terminal parity
72/72 exact verdict and reason; 10/10 exit criteria
0 mismatches between Node and Python verifiers
Production conformance cases
42/42 cases across 28 required classes
Production independent parity for MutMem V2 protocol
Canary explicit marker lane
60/60 marked; 0/60 clean
100.0% detection; 0.0% false positive for explicit markers
These numbers come from MutMem V2’s reproducibility and conformance evidence table, which treats each vector or case as a verification unit rather than a utility benchmark.
They show that MutMem V2’s portable protocol, verifiers, and conformance corpus align exactly on terminal verdicts and reasons, and that explicit marker traversal is fully detected within its Canary scope.
BENCHMARK
These numbers come from MutMem V2’s reproducibility and conformance evidence table, which treats each vector or case as a verification unit rather than a utility benchmark. They show that MutMem V2’s portable protocol, verifiers, and conformance corpus align exactly on terminal verdicts and reasons, and that explicit marker traversal is fully detected within its Canary scope.
BENCHMARK
Verification units across MutMem V2 protocol, verifiers, production corpus, and Canary lane.
KEY INSIGHT
MutMem V2 reports 72/72 exact verdict and reason parity across independent Node and Python implementations, plus 10/10 exit criteria matches.
This is surprising because complex cryptographic and structural protocols often diverge in edge cases, yet MutMem V2’s closed failure vocabulary forces identical reasoning across languages.
WHY IT MATTERS
MutMem V2 unlocks a way to ship persistent agent memory with a portable, independently checkable integrity and authorization contract rather than an opaque database.
Builders can now provide reproducible, cryptographically bound evidence for recall and mutation behavior, enabling third party reviewers to verify agent memory claims without reimplementing the runtime.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Answers use this explainer on Memory Papers.
Checking…