DecentMem mapping (Hao et al. 2026)

Document classification: Public-facing strategic supplement.

Date: 2026-06-15.

Status: Reference material. Not a constraint. The moonshot synthesis (docs/strategy/moonshot-synthesis.md) and the seven §2 properties in ROADMAP.md remain the authoritative anchors. Where DecentMem and the moonshot disagree, the moonshot wins.

Purpose. Map ai-memory’s substrate primitives to the decentralized-memory vocabulary of Self-Evolving Multi-Agent Systems via Decentralized Memory (Hao, Long, Zhao, arXiv:2605.22721) for readers familiar with the multi-agent-systems literature. This document does not derive the substrate’s properties from DecentMem — those derive from the moonshot. DecentMem serves as a retrospective organizing lens.

1. Executive position

ai-memory already implements the structural primitives DecentMem’s per-agent decentralized memory requires — store, retrieve, consolidate, score, expire, per-agent isolation — and operates one layer below DecentMem’s contribution. DecentMem is a multi-agent-systems orchestration strategy (per-agent dual pools, online utility reweighting); ai-memory is the substrate such an orchestrator runs on. DecentMem’s value here is academic legibility for the MAS-memory literature, not architectural authority. The full assessment is recorded in issue #1704.

2. DecentMem in one paragraph

DecentMem gives each agent in a self-evolving multi-agent system its own dual-pool memory: an exploitation pool of consolidated past trajectories and an exploration pool of LLM-generated candidate memories for unseen contexts, reweighted online by an LLM-as-a-judge contextual bandit (claimed O(log T) regret, global reachability). It is decentralized: each agent keeps its own memory with no central coordination, argued to preserve agent diversity (the paper’s thesis is that a centralized shared memory homogenizes a collective and collapses its diversity) and to reduce coordination/privacy cost. DecentMem does not address attestation, endpoint residency, separation-of-powers reflection, or cryptographic non-repudiation; its evaluation is end-to-end MAS task accuracy inside AutoGen / DyLAN / AgentNet, a metric that is only measurable against a full agent collective, not against a memory substrate.

3. Side-by-side mapping

Descriptive only. Code anchors are current as of v0.7.1.

3.1 Dual-pool memory

DecentMem primitive	DecentMem definition	ai-memory realization	Code anchors
Exploitation pool	Consolidated past trajectories the agent reuses	`MemoryKind::Reflection` + the Observation→Reflection→Skill curator pipeline (consolidates co-recalled observations into typed reflections with `reflects_on` provenance). Note: ai-memory consolidates co-recalled, not task-succeeded, trajectories — the success signal is an orchestration concern.	`src/curator/reflection_pass.rs`, `src/storage/reflect.rs`, `memory_consolidate`, `MemoryKind` enum
Exploration pool	LLM-generated speculative candidates for unseen contexts	Intentionally absent. Every ai-memory write originates from a real observation/ingest/reflection over real data; there is no proactive speculative-candidate generator (`src/synthesis/mod.rs` is reactive dedup only). Generating memories the agent never observed cuts against §2.5 (attestation — a speculative memory has no attestable provenance) and §4 (cognitive artifacts of engagement, not bare propositions). This is agent-orchestration behavior, scope-out per §16.	(no production path; `src/synthesis/mod.rs:141`)
Per-agent isolation	Each agent’s pool is private; no shared corpus	`MemoryScope::Private` (default-deny) + `is_visible_to_caller` + per-agent `metadata.agent_id`. An operator can deploy ai-memory in a fully per-agent-isolated posture today.	`src/models/namespace.rs:124`, `src/visibility.rs:46`

3.2 Online reweighting / feedback

DecentMem primitive	DecentMem definition	ai-memory realization	Code anchors
Utility scoring	LLM-judge contextual bandit reweights memories by downstream task success	Static substrate primitives an orchestrator would compose: 6-factor recall blend, Form-5 confidence calibration, cross-encoder reranker with score floor. None is an online success-driven reweighter — the bandit control loop is agent-layer.	`src/storage/mod.rs:2949-2954` (blend), `src/confidence/`, `src/reranker.rs`
Stage-wise feedback	Per-stage success signal feeds the reweighter	The `recall_observations` ledger records the input half of this loop: `record_recall` logs each candidate; `mark_consumed` flips `consumed` + `consumed_by_memory_id` when a recalled memory is cited downstream. The loop is currently open — the consumed signal is recorded telemetry; no production scoring path reads it back (the only reader is the `memory_recall_observations` inspection tool).	`src/observations/mod.rs:65,108,162`

Scope firewall. Closing the recall_observations loop — feeding the consumed signal back into recall scoring — is tracked as a shadow-first v0.9 enhancement (#1706), gated on a v0.8.0 ledger backend-parity + integrity prerequisite (#1705); the live ranking change is a conditional follow-up (#1707). This mapping describes current state only and proposes no design.

3.3 Decentralized coordination

DecentMem mechanism	ai-memory realization	Code anchors / ROADMAP
No central coordination	v0.7.0 baseline is shared-nothing per-agent SQLite-1:1 + CRDT P2P federation (newer-wins merge with deterministic id-tiebreak) — not the centralized shared repository DecentMem argues against	`src/storage/mod.rs:7309` (`insert_if_newer`), ROADMAP §11.4 Pillar-4 (`ROADMAP.md:608-610`)
Diversity preservation	The §11.4 v0.8 module model introduces an opt-in shared consolidated Postgres+AGE corpus, and the ROADMAP independently names the diversity-collapse risk DecentMem describes	ROADMAP §5 (`ROADMAP.md:183`), §11.4 Pillar-4

4. What ai-memory adds beyond DecentMem

ai-memory property	DecentMem coverage	Substrate’s structural answer
§2.1 Endpoint-resident	None — DecentMem is agnostic to deployment topology	Rust core + SQLite default + LLVM-portable; mobile cross-compile gate (#1068)
§2.5 Attested with cryptographic non-repudiation	None — DecentMem does not address audit or attestation	V-4 `signed_events` cross-row hash chain, per-agent Ed25519 attestation
§2.3 Stoppable without silent corruption	None	Typed refusals, `permissions.mode = enforce` fail-CLOSED, attested checkpoints
§2.6 Bias-displaced through separation-of-powers	Partial — DecentMem’s decentralization preserves diversity but has no reflection-boundary primitive	LLM-agnostic reflection boundary; §5 decorrelation enforcement (committed v0.8/v0.9)

DecentMem’s decentralization thesis and ai-memory’s §2.6/§5 decorrelation commitment are the same concern from different angles (a shared corpus homogenizes the agents drawing from it). DecentMem is a second independent corroboration of that axis — after the DeepMind paper (#1698) — but it changes no commitment: §5 is already committed.

5. The DecentMem-named direction ai-memory is closing

Recall-usage feedback. DecentMem’s stage-wise reweighting (favor memories that contributed to downstream success) is the concept of feedback-driven recall-utility weighting. ai-memory already records the input signal (recall_observations.consumed) and is closing the loop shadow-first (#1706), correctly de-scoped from DecentMem’s online per-stage LLM-judge to an offline calibration sweep that respects the substrate’s ~35 ms recall p95 budget. This is a direction in progress, not a closed gap.

6. Honest limits of the mapping

DecentMem operates at the agent-orchestration layer; ai-memory is the substrate below it. Two of DecentMem’s four mechanisms (the exploration pool, the online bandit reweighting) are agent-runtime behaviors ai-memory deliberately does not implement (§4 / §16 scope test). The mapping describes structural equivalences for the primitives an orchestrator would compose, not feature parity.
DecentMem’s anti-centralization critique largely misses ai-memory’s architecture. ai-memory’s baseline is shared-nothing federation, not a central repository — so the paper’s headline empirical win is against an architecture ai-memory does not use. The diversity-collapse concern lands only on the opt-in v0.8 shared-corpus module, which §5 already addresses.
The headline numbers are not substrate-comparable. DecentMem reports MAS task-accuracy (and the abstract’s “+23.8% / −49%” are single best-cell figures; the body averages are ~8.6% accuracy / ~43% tokens). These measure a full agent collective, not a memory store; they cannot be claimed or refuted by ai-memory’s recall-fidelity metrics (LongMemEval, §9.5).
The O(log T) regret theory is bandit-framing-specific. A storage substrate has no arm, round, or reward, so the regret bound does not apply to ai-memory.

7. Disposition

This document is reference material. It does not commit the substrate to any DecentMem-specific implementation. The seven properties in ROADMAP §2 are the authoritative test. The §3 scope test is controlling. The moonshot synthesis is the North Star.

If a future revision or successor work changes the taxonomy, this document will be revised to track. The substrate’s commitments will not change in response to framework drift.

References

Hao, G., Long, Y., & Zhao, Z. (2026). Self-Evolving Multi-Agent Systems via Decentralized Memory. arXiv:2605.22721.

Revision history:

2026-06-15 (initial): created as reference supplement, derived from the 5-agent adversarial assessment (#1704) and the 7-agent execution-design panel. No substrate code changes. No commitments added or modified. No §2 properties changed. No §3 scope-test modifications.

End of document.