AtMem 2.3.8 — Governed context and one-command Hermes setup
AtMem 2.3.8 adds a source-linked context engine and a reversible Hermes installation path while preserving AtMem's role as the memory and governance authority. It forms typed evidence from immutable source episodes, retrieves different evidence shapes for facts, transitions, procedures and action constraints, expands bounded source neighbourhoods, and reports whether the assembled context is sufficient, partial, conflicted or unsupported.
Install or upgrade
Back up AtMem Home and stop writers before upgrading. In the Python environment used by the agent:
python -m pip install --upgrade "atmem==2.3.8"
atmem --version
atmem init
atmem statusAtMem installs the compatible companions atmem-atbot==0.1.0 and
atflows==0.1.4b3. AtFlows remains optional at runtime. Long-running dashboards
and agents must be restarted after upgrade.
For OpenClaw, use the Python-owned installer so the matching
openclaw-memory-atmem@2.3.8 bridge, saved configuration and verification steps
stay aligned:
atmem openclaw upgrade
atmem control verify --probeDo not install the npm bridge directly.
Hermes setup
The safe default creates an isolated Hermes memory scope and starts in shadow mode. It does not share OpenClaw or another agent's memory and does not inject AtMem context into the model:
atmem install hermes --yes
atmem hermes status --jsonTo make sharing and active recall explicit:
atmem install hermes --memory shared --activate --verify-turn --yes
atmem hermes status --jsonThe installer discovers Hermes, records its previous provider, installs the native AtMem provider, imports supported existing memory as reviewable proposals rather than silently admitting it, and can verify one temporary recall lifecycle. Restore is reversible and refuses to overwrite later user edits:
atmem restore hermes --yesFor bounded native Hermes model/tool observation, install or retain AtFlows 0.1.4b3 and use its guided Hermes connection. Memory works without AtFlows.
Context and retrieval changes
- Immutable source episodes now form typed facts, rules, procedures, state transitions, gotchas, premises and raw-state ranges with source links.
- Information-need routing uses exact, lexical, semantic, temporal and graph nominations without granting any candidate authority by similarity alone.
- Bounded evidence neighbourhoods keep the trigger, condition, action, outcome and adjacent identity details together.
- Context packing prioritizes complementary coverage and keeps addressed-action evidence contiguous, including exact dates, identifiers, recipients and source-backed email addresses.
- Sufficiency receipts distinguish complete evidence from missing obligations. The Dolphin adapter can block before model or tool invocation and names the missing requirement; AtMem itself remains a context provider rather than an action executor.
- Agent scopes remain isolated by default. Explicit shared-memory setup reuses a governed scope; it does not merge all agents' data into one unscoped table.
Bounded evaluation evidence
The release-candidate code was measured on frozen development samples with the same reader and judge routes for each compared arm. These are not official or statistically representative leaderboard scores.
| Development surface | AtMem | Matched comparison |
|---|---|---|
| LongMemEval-V2, 23/451 questions | 5/23 | AgentRunbook-R 4/23; Mem0 1/23 |
| DolphinBench, 30/600 tasks | 9/30 | Mem0 7/30 |
| DolphinBench checks | 32/97 | Mem0 27/97 |
| AGMI audit chain only | 8/9 | Whole-store rollback remains undetectable |
| AGMI with external checkpoint | 9/9 | All nine pinned attacks detected |
All system failures remained in the denominators. The final Dolphin arms had zero system failures. The Mem0 Dolphin environment lacked its optional spaCy and fastembed extras. Read the exact hashes, controls and limitations in the frozen 5% report.
Migration, compatibility and limits
- No canonical-memory database migration is required solely for this upgrade. Existing encrypted stores remain authoritative; V3 context views are derived and can be rebuilt from retained sources.
- Encrypted household activation still requires a working SQLCipher runtime. Stop all writers and retain a backup before any explicit encryption migration.
- Hermes setup was verified end to end on macOS with Hermes Agent 0.21.5. Linux uses the same local contract; Windows Hermes uses the documented WSL route. Native Windows credential-ACL parity is not claimed.
- The OpenClaw declaration matrix covers the minimum
2026.7.1-2, the pinned2026.8.1, and reviewed 2026.9.x fixtures through2026.9.9. Unlisted host versions remain fail-closed until their hook surface is reviewed. - Shadow mode withholds context by design. Active injection requires
--activate; cross-agent sharing requires--memory shared. - AtFlows 0.1.4b3 observes bounded native Hermes metadata, not every possible host event. Generic proxy/OTLP redaction remains an AtFlows release concern.
- Audit-chain-only storage cannot detect a valid whole-store snapshot rollback. The 9/9 AGMI result requires a trusted checkpoint outside the attacker-controlled store directory.
- The 5% evaluations do not justify a full-benchmark, leaderboard or universal superiority claim. Full held-out qualification remains future work.
The detailed architecture and evidence changes from 2.3.7 are recorded in the 2.3.7 to 2.3.8 technical report.