AtMem 2.3.10 — Bounded evidence discovery

AtMem 2.3.10 promotes the tested 2.3.10b1 product to a stable release. It gives explicitly configured hosts scoped, read-only evidence discovery, while preserving ordinary deterministic memory, Code, integrity and governance defaults. No Laya model, new training or checked-experience learning is included.

What users get

• Bounded memory_search and memory_inspect tools with original quotations, source provenance and live scope checks, rather than an unbounded context dump.
• Administrator-issued Hermes discovery leases with persisted limits and revocation.
• Optional, explicitly selected semantic nominations and AtBot source-state interpretation. No automatic model download or hidden remote provider activation.
• Encrypted derived-cache cleanup, stale-generation rejection and source-first neighbor/head lookup with canonical revalidation.
• The beta's native Windows ACL and SQLite capability qualification fixes.
• Existing shared Code structure/experience, project portability, dashboard, OpenClaw and framework integrations remain available with their existing limits.

Discovery remains opt-in SDK/host functionality. Installing the stable release does not enable it in every agent CLI or change ordinary retrieval defaults. Start with the evidence discovery guide or the Code integration guide.

Install or upgrade

Use Python 3.10–3.13 in an isolated environment:

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

Optional encrypted discovery or shared coding memory:

python -m pip install --upgrade "atmem[encrypted,context-retrieval]==2.3.10"
# Or, for the coding flavor:
python -m pip install --upgrade "atmem[code]==2.3.10"

These pip commands also work in Windows PowerShell after activating your virtual environment. No --pre flag, GPU, Torch or API key is needed for the base installation or model-free encrypted example. On Linux, install the SQLCipher headers/build tools required by the encrypted extra; follow the platform guide.

Do not install AtBot separately. For an existing OpenClaw connection, upgrade Python first, then run atmem openclaw upgrade. Use atmem openclaw install for a new connection, not a direct npm installation. Existing host modes remain unchanged; host opt-in and authorization are still required for discovery. The Python-owned installer selects openclaw-memory-atmem@2.3.10 exactly.

Existing data and rollback

Stop writers and make a verified copy of your Home/database, keys and evidence before upgrading. Existing canonical data uses the normal migration path. No embedding rebuild is required for installations that never enabled semantic discovery. Optional semantic caches remain disposable derived data, not authority. Restart long-running clients and managed companions after upgrading.

For rollback, use a separate older-version environment with its preserved pre-upgrade copy. Do not perform an in-place downgrade of a Home already changed by newer code. No new schema or retrieval behavior is introduced by the stable version promotion itself.

Windows uses the authenticated-owner delegated-context path. The POSIX-only origin-metadata-absent exception stays refused on Windows; unsafe permissions, reparse paths and unverifiable files do not become authorized by an upgrade.

Benchmark provenance

The published 600-task DolphinBench dataset records AtMem 2.3.10b1, not a newly executed stable-version campaign: 390/600 tasks passed (65.0%), with 85.78-second median recorded task latency. Its original ZIP and hashes remain unchanged. See DolphinBench PR #10 for the self-tested submission and documented configuration/recovery provenance. Maintainer acceptance is separate from package publication.

The stable promotion qualification records the version-only mapping and applicable gates. AtBot/AtFlows dependency pins, retrieval logic, model prompts and defaults are unchanged. Version metadata changes mean the beta and stable wheels are not byte-identical archives.

Website publication

Public documentation is updated with the release. The private AtMem.ai source-pin refresh requires a separate owner-reviewed PR; only owner-merged private main may subsequently be deployed to Firebase. Package availability, website merge and live deployment are separate publication states.