Skip to content

v2.3.9b1

Pre-release
Pre-release

Choose a tag to compare

View all tags
@github-actions github-actions released this 11 Oct 01:18
· 2 commits to main since this release
v2.3.9b1
939a031

AtMem 2.3.9b1 — Optional local Laya memory formation

AtMem 2.3.9b1 adds an opt-in, locally runnable Laya System One profile for
bounded memory-formation decisions. It classifies the permitted operation,
memory class, evidence support, authorized target and retrieval usefulness;
AtMem still owns authority, state mutation, post-write validation, fallback,
review and deletion. Existing deterministic behavior remains the default.

The release also publishes a 12,000-scenario/60,000-decision synthetic dataset
and the fine-tuned model in the atmem Hugging Face organization:

Install or upgrade

Back up AtMem Home and stop long-running writers before upgrading. The normal
installation does not install Laya or download a model:

python -m pip install --upgrade --pre "atmem==2.3.9b1"
atmem --version
atmem init
atmem status

This installs exact companions atmem-atbot==0.1.1b1 and
atflows==0.1.4b3. Restart long-running agents and dashboards after upgrade.

For OpenClaw, let AtMem install the matching prerelease bridge
openclaw-memory-atmem@2.3.9-beta.1:

atmem openclaw upgrade
atmem control verify --probe

Do not install the npm bridge directly.

Opt in to local Laya formation

Install the optional runtime explicitly:

python -m pip install --upgrade --pre "atmem[laya-formation]==2.3.9b1"
atmem formation preview --download
atmem formation setup --download --device auto --yes
atmem formation doctor
atmem formation activate --yes
atmem formation status

preview changes nothing. setup anonymously downloads the exact public model
revision, verifies every artifact and calibration binding, and stages an
inactive profile. Only activate --yes changes the selected formation profile.
The model proposes finite choices; it cannot widen evidence, scope or mutation
authority. If the model rejects, times out, overflows or abstains, AtMem uses the
declared deterministic fallback/review path.

To return to the byte-identical staged state saved before activation:

atmem formation rollback --yes
atmem formation status

Bounded AtBot escalation is a separate opt-in. Add --escalation during
preview/setup, then configure an AtBot provider explicitly. Setup does not
enable remote egress: the profile starts with one-call/no-retry limits, zero
allowed remote cost and egress denied until the operator changes provider
policy.

User-visible benefit and measured evidence

On the frozen synthetic AtMem Formation benchmark, fine-tuned Laya produced the
exact canonical state for 80/80 scenario clusters, versus 34/80 for current
deterministic AtMem and 29/80 for unmodified Laya. It had zero safety violations
versus 19 and 3 respectively. Ten thousand paired cluster bootstrap resamples
put the exact-state improvement over deterministic AtMem at +0.575, with a 95%
interval of +0.468 to +0.684.

The strict store-then-retrieve workflow passed 80/80 clusters with fine-tuned
Laya versus 7/80 with deterministic AtMem. The improvement was in deciding what
state to store, not in ranking: both arms had retrieval MRR@5 0.3798 and
Recall@5/10 1.0 over 61 positive fixed-pool queries. On an Apple-silicon CPU,
fine-tuned formation took a median 272 ms per finite choice versus 2.3 ms for
the deterministic rules. The bounded-escalation arm made zero provider calls in
this run.

These are synthetic results, not production or general customer evidence. Jev
exposed no concrete immutable model identity, and the pinned Qwen deployment
failed before inference, so every formation, retrieval and combined comparison
against Jev or Qwen is inconclusive. No intended-scale candidate run exists for
both LoCoMo and LongMemEval/store-to-retrieval. See the
checksum-bound evaluation report.

Compatibility

The base wheel passed clean source-isolated installs on Python 3.10, 3.11, 3.12
and 3.13. The exact optional model and candidate wheel completed real bounded
inference on:

Host CPU Accelerator
macOS 26.5.2, Apple M2, Python 3.12.13 passed MPS passed
Windows 11 build 26200, Python 3.12.6 passed CUDA passed on Quadro P2000
Ubuntu 22.04, Python 3.11.13 passed CUDA passed on RTX 2070 Super

Explicit CUDA on macOS and MPS on Windows/Linux fail before activation with an
actionable diagnostic. The full optional model was tested on Apple-silicon
macOS, not Intel macOS. A normal base install remains model-free and works
without PyTorch or Laya.

Migration, compatibility and limits

  • Upgrading from published AtMem 2.3.8 requires no canonical-memory migration.
    A persisted 2.3.8 record remained recallable after candidate upgrade and
    exact rollback.
  • Existing CLI, MCP, HTTP, AtBot protocol, OpenClaw and AtFlows contracts remain
    compatible; formation status is additive in CLI, MCP and dashboard surfaces.
  • OpenClaw hook-context compatibility is fixture-reviewed through 2026.10.1;
    unlisted host versions still fail closed rather than being silently accepted.
  • Laya 0.4.2, the model revision, questions, tokenizer and calibration are
    pinned and checksum-bound. The model export is about 800 MB; optional runtime
    dependencies require additional disk and memory.
  • The shipped calibrated questions use three to five choices. Laya may warn
    about an unused 11+ choice calibration bucket during load; that bucket is not
    used by this profile, and AtMem installs and verifies its own bound 3–5-choice
    calibration before inference.
  • Model decisions remain proposals. AtMem reloads and revalidates canonical
    state before disposition and never grants a model direct storage authority.
  • Synthetic perfect accuracy can reflect the controlled generator and does not
    establish behavior on real customer language, all adversarial inputs or
    every memory domain.
  • The release is a beta. The model profile is inactive by default, never
    downloads during ordinary upgrade, and is fully reversible.
Assets 5
Loading