AtMem From remembering relevant text to assembling usable evidence
AtMem 2.3.8 focuses on that gap: turning retained history into evidence an agent can use, inspect, and trust within a defined scope.
What changed from 2.3.7?
Version 2.3.7 already had governed memory records, scope checks, audit trails, context receipts, and recovery for supported interrupted workflows. It also had lexical, graph, and semantic retrieval capabilities.
Version 2.3.8 adds a new context engine. It preserves immutable source episodes, builds typed views over exact source ranges, and breaks a request into explicit evidence needs.
Retrieval finds supporting ranges. Packing keeps complementary facts and complete action episodes together. Before delivery, a separate governance layer reloads the selected evidence and checks its scope, generation, lifecycle, and size.
The practical question is:
Did the evidence required for the task actually survive into the context the agent receives?
What is novel?
The contribution is the connection between evidence needs, source preservation, packing, and governed delivery.
Typed pools are not new; AgentRunbook-R already uses them. Audit chains and source attribution also have substantial prior art.
AtMem connects these ideas through one inspectable pipeline, with receipts showing what was covered, missing, conflicting, or withheld.
For a request to email a review outcome, this can mean preserving the trigger, result, restriction, and recipient evidence together. If required evidence is missing, an integration can block before invoking a model or tool.
What did we measure?
On frozen development samples, the AtMem 2.3.8 candidate achieved:
- LongMemEval-V2: 5/23 correct
- AgentRunbook-R: 4/23
- Mem0 OSS: 1/23
- DolphinBench: 9/30 tasks passed
- Mem0 OSS on DolphinBench: 7/30
- DolphinBench checks: 32/97 versus Mem0’s 27/97
These are small development samples. They are neither complete benchmark results nor statistically significant wins.
AgentRunbook-R was not evaluated on DolphinBench. The local configurations used hash embeddings. Mem0 ran with inference disabled, and its DolphinBench arm lacked optional ranking components.
The results therefore do not represent the strongest published configurations of either competitor. LongMemEval also retained six AtMem system failures in the denominator.
The evidence-preserving changes provide plausible explanations for the observed results. Component ablations and untouched evaluation tasks are still required to determine how much each mechanism contributes.
Memory integrity also improved
AGMI exposed a concrete gap: verifying an audit chain did not establish that the live memory records still matched the history committed to that chain.
AtMem 2.3.8 adds whole-record commitments and reconciliation between protected audit events and live records.
In our reproduction of the unchanged AGMI 0.6.3 harness:
- The audit chain alone reported 8/9 attack classes
- The audit chain with a trusted external checkpoint reported 9/9
The external checkpoint is required to detect whole-store snapshot rollback because a rolled-back database cannot independently prove that a newer state once existed.
Independent upstream reproduction remains pending. These are audit-detection results; they do not imply that every read path automatically blocks tampered memory.
Why this matters
AtMem’s long-term case is architectural: useful memory whose source, authority, scope, completeness, delivery, and integrity can be investigated.
The goal is memory that can explain:
- What evidence was retained
- What the request required
- Which sources supported it
- What was lost during formation or packing
- Why delivery was allowed or blocked
- Whether live records still match protected history
The next step is stronger baseline configurations, held-out evaluation, symmetric repetitions, and component-level ablations—backed by the same explicit evidence.
Release and installation:
https://github.com/aetna000/atmem/blob/v2.3.8/docs/releases/v2.3.8.md
Results and limitations:
