AtMem 2.3.10b1 — Bounded evidence discovery

This beta packages opt-in, source-backed investigative retrieval for host builders while retaining the stable deterministic memory and Code paths. It is the candidate line for the planned full benchmark, not a claim that the full benchmark has already passed. Publication and platform qualification are tracked separately in the release preparation report.

What users get

• Read-only memory_search and memory_inspect tools with original quotations, provenance, live scope checks and bounded session budgets.
• Administrator-issued Hermes discovery leases whose limits and revocation survive restarts, independent of existing automatic capture/injection.
• Explicit SDK selection of lexical/semantic nominations and quoted AtBot source-state interpretations. No hidden provider activation or model download.
• Source-first indexed neighbor/head lookup without changing evidence selection, source order, lifecycle checks or sufficiency rules.
• Registered encrypted source-part cache cleanup and rejection of stale/deleted generations; missing cache verification cannot be labelled successful deletion.
• Native Windows ACL protection and verification for delegated-provider signing, configuration and HMAC files, instead of unusable POSIX-mode checks.

AtMem Code structure/experience, portability, ordinary memory, audit/governance, OpenClaw and framework integration defaults remain in place. The new discovery workflow is opt-in SDK/host functionality, not automatic activation in every CLI.

Install or upgrade

Use Python 3.10–3.13 in an isolated environment. Once the beta is published:

python -m pip install --upgrade --pre atmem==2.3.10b1
atmem --version
atmem status --json

For the optional encrypted discovery example:

python -m pip install --upgrade --pre "atmem[encrypted]==2.3.10b1"

For the explicitly configured native vector backend:

python -m pip install --upgrade --pre "atmem[encrypted,context-retrieval]==2.3.10b1"

Windows PowerShell accepts these same pip commands after activating .\.venv\Scripts\Activate.ps1. No GPU, Torch or remote API key is needed for the base install or the model-free example. Follow the step-by-step evidence discovery guide.

Compatibility

The Python-owned OpenClaw installer pins openclaw-memory-atmem@2.3.10-beta.1. Upgrade the Python package first, then use:

atmem openclaw upgrade

Use atmem openclaw install for a new connection. Do not replace the managed installation with a direct npm install. Existing host modes are preserved by upgrade; adding this beta does not itself enable investigative discovery.

Windows uses the normal authenticated-owner path for delegated context. The origin-metadata-absent isolated-CLI exception relies on POSIX permissions and is explicitly refused on Windows; this does not disable ordinary owner-bound OpenClaw memory. See the identity contract. Provider setup uses owner-only Windows ACLs; reads reject ordinary additional users, broad grants, reparse points and unverifiable permissions. SYSTEM and local Administrators remain trusted OS authorities. Unsafe existing provider files are not silently accepted or repermissioned during a read; repair their permissions deliberately or create a new isolated provider configuration.

Existing data, migration and rollback

Stop writers and make a verified Home/database copy before upgrading. Keep keys and evidence with the copy. Existing canonical data follows the normal migration path; installations that never enabled semantic discovery do not need embeddings or a semantic-cache rebuild. New optional caches register in the existing semantic registry and remain disposable derived data, not memory authority.

Restart long-running clients/managed companions after upgrading and confirm the reported versions and host state. For rollback, use a separate 2.3.9 environment with the preserved pre-beta copy. Do not promise a safe in-place downgrade of a Home already modified by newer code.

Beta-to-stable promotion

Freeze the exact beta wheels, source revision, dependency/environment manifests, configuration and model identities before the 600-task campaign. Preserve all attempts. A version-metadata-only promotion to 2.3.10 can reuse those results after installed-package, migration and equivalence checks; identify the tested beta and stable mapping explicitly. Changes to behavior, prompts, models, dependencies or defaults require affected revalidation and potentially a new full run. Never move a published beta tag or call stable binaries byte-identical when their version metadata changed.

The current gate is release-quality preparation, not a performance claim. Research acceptance targets, the full600 run and matched Mem0 superiority remain separate pending work. No Laya model/training or checked-experience learning feature is included in this beta.