AtMem 2.3.7 → 2.3.8 technical report
Summary
AtMem 2.3.7 concentrated on durable execution continuity: governed operation identities, receipts, restart recovery and inspection. Version 2.3.8 keeps that surface and changes how agent memory is formed, selected and delivered. The new engine treats memory as source-linked evidence with explicit scope and sufficiency, then adds a managed Hermes host path and benchmark instrumentation that measures where evidence is lost.
Architecture evolution
| Area | 2.3.7 | 2.3.8 |
|---|---|---|
| Source model | Canonical records and retained evidence | Immutable source episodes plus exact source ranges and derived typed units |
| Formation | General memory extraction | Facts, rules, procedures, transitions, gotchas, premises, entities and raw state with coverage/loss receipts |
| Retrieval | Primarily ranked candidates | Information-need routing plus exact, lexical, semantic, temporal and graph nomination |
| Context | Ranked bounded output | Obligation-first complementary packing with bounded source neighbourhoods |
| Decision | Relevance and policy checks | sufficient, partial, conflicted or unsupported, with named evidence obligations |
| Multi-agent boundary | Governed subject/agent/workspace scope | Same authority boundary, exposed as isolated-by-default or explicitly shared host setup |
| Hermes | No managed first-class setup | Preview/apply/status/restore installer, native provider, shadow/active modes and optional verified turn |
| Integrity | Local audit-chain protection | Stronger chain binding and optional external checkpoint for whole-store rollback detection |
| Evaluation | Product/continuity fixtures | Frozen LongMemEval-V2, DolphinBench and AGMI adapters with installed-artifact, route and failure attribution receipts |
Retrieval-quality work
The retrieval changes were driven by missing evidence rather than by increasing
top_k. Formation preserves exact values and source ranges. The planner routes
current-state, historical transition, ordered-procedure, exception, comparison
and action questions differently. Retrieval then expands only bounded adjacent
ranges and packing groups evidence by requested obligation and source episode.
The final action path adds three general corrections found during the frozen Dolphin development work:
- Exact dates and identifiers receive priority nominations without evaluator labels or expected answers.
- A late-selected action source can still contribute its triggering statement, outcome and confidence/value tail as one contiguous episode.
- Indirect recipients such as an integration contact or the usual external copy recipient retain source-backed adjacent addresses rather than forcing the reader to infer or ask the user again.
The benchmark adapters consume the same context API as product integrations. They cannot mutate AtMem during scoring or read evaluator requirements. Removal controls classify a safe block only when the named requirement is missing, the model was not invoked and no tool call occurred; timeout, parse error, provider error and silent no-call never count as a safety pass.
Evaluation trajectory
On the frozen 30-task Dolphin sample, the earliest recorded AtMem arm passed 2 tasks and 15 of 97 checks. After source-neighbour, temporal and identity routing work, the final candidate passed 9 tasks and 32 checks with zero system failures. The unchanged matched Mem0 arm passed 7 tasks and 27 checks. This is a development trajectory, not a leaderboard result.
On the frozen 23-question LongMemEval-V2 sample, AtMem answered 5 questions, AgentRunbook-R answered 4 and Mem0 answered 1. A verified-evidence control answered 19, demonstrating that reader use and remaining evidence delivery are still the dominant improvement opportunity. Six AtMem system failures remained in that run and are included in the denominator.
AGMI detected 8/9 attacks with the audit chain alone and 9/9 when the checkpoint file was held outside the attacker-controlled store directory. The difference is a security boundary, not a scoring trick: an attacker who can replace the entire store can also restore a valid old chain unless trusted state exists outside that store.
Product and operational boundaries
- AtMem remains the context and governance provider. Host adapters decide when to call a model or tool and may enforce AtMem's sufficiency receipt.
- Agent memory is isolated by default. Sharing is explicit and scope-based; telemetry remains separately bucketed by host/agent/session in AtFlows.
- Source text, derived indexes and telemetry are not collapsed into one search table. Retrieval uses bounded derived views while canonical evidence remains reconstructible and governed.
- Heavy benchmark datasets and transient model artifacts stay outside the source repository. Product code, protocols, compact receipts and reports stay versioned in the repository.
- The 5% samples are useful release-development evidence but cannot establish a full benchmark result, statistical significance or a general claim that AtMem beats every memory system.
Reproducibility
The frozen IDs, provider/model revisions, installed artifact hashes, evaluator
settings, costs, system failures and result hashes are recorded in
benchmarks/retrieval_quality/baselines/five-percent-results-20261010.json and
the accompanying report. Historical beta artifacts keep their original
2.3.8b6 identity; stable metadata does not relabel those measurements.