Repository navigation
AtMem 2.3.15 — Governed memory tools and measured evaluation
AtMem 2.3.15 is a stable package release. It brings scoped document retrieval and
host-neutral MCP, inspectable action outcomes and explicitly invoked memory-policy
studies to stable users while preserving ordinary memory/context defaults.
The practical change is accountability: an operator can inspect what evidence was
retrieved, what action was attempted, what independent readback established, and
whether a proposed memory policy actually helped. Retrieval or a generated lesson
is never reported as proof of learning. No reliable learning gain is claimed.
Install or upgrade
Use Python 3.10–3.13 in a virtual environment:
python -m pip install --upgrade atmem==2.3.15
python -m pip check
atmem --version
atmem statusFor SQLCipher and recipient-crypto paths:
python -m pip install --upgrade "atmem[code]==2.3.15"On Linux, building the SQLCipher extra may require the documented native SQLCipher
headers/compiler. Installation does not start paid inference. Start local services
with atmem init --no-open; use --home PATH for an isolated installation.
Exact compatibility: AtMem 2.3.15, openclaw-memory-atmem@2.3.15 (npm latest),
unchanged atmem-atbot==0.1.1 and atflows==0.1.8. Install/upgrade the bridge through
atmem openclaw upgrade, then inspect atmem openclaw status. Historical b1/b2
research identities remain in the evidence; b2 was never a published package.
What changes for users
- Owner-scoped existing-document retrievers and an encrypted local document
provider return original cited passages and source revisions. Date/jurisdiction
filters and post-query authorization checks do not establish factual or legal truth. - A separate authenticated local stdio MCP/CLI surface supports scoped retrieval,
read/report/propose operations and registered guarded proxy handlers. Connecting
MCP does not control unrelated host tools or grant model arguments owner authority. - Per-action outcomes distinguish reads, writes and independent backend readback.
Missing, disputed or incomplete evidence remains unknown. Dashboard, CLI and MCP
use the same authority state; exports require recipient/purpose-bound consent. - An owner-only SDK and
atmem-memory-studycommand freeze study inputs, policy,
verifier and budgets, retain exact exchanges encrypted, and preserve failed and
inconclusive experiments. No automatic activation, model training or background
optimization is introduced. - The source-only installed-package comparison quotes historical evidence
separately from the current task. This fixes a research adapter defect that let
old questions displace the current question. Malformed replies fail without paid
retries, and tool observations are captured before another model call. - Windows binary migration/restore and long checkpoint paths retain their fixes.
macOS CI is pinned to the supported ARM64 image to avoid floating-image drift;
applicable installed and persisted-upgrade gates remain required.
Opt-in, migration and rollback
Ordinary memory, recall and existing MCP defaults are retained. Importing feedback
does not activate it. Governed RAG/MCP requires owner setup, explicit scoped grants
and an existing evidence store. Studies require an explicit paid/public-data flag
before network inference; their reports cannot issue activation authority.
Back up the complete AtMem Home, canonical databases, encryption policy, keys,
accounts and checkpoints before enabling feedback. See
storage and backup. Enabling feedback opts that
existing evidence container into format 4. Older readers must refuse it; unused
containers/native memory retain their behavior. There is no in-place downgrade:
restore the complete pre-opt-in backup to a separate Home and use supported
adoption/reconciliation. Restore reconciliation quarantines old policy/grants;
it does not reactivate authority. This release adds no new canonical memory schema.
Evidence and limits
The original four-arm notes experiment failed to improve results: strict
no-memory 7/20 versus revised notes 5/20. Generic notes were not an original-AtMem
baseline. A separate final check exercised installed original AtMem 2.3.14 and the
retained candidate through actual encrypted Context Engine V3. On six independently
authored public database tasks, all three conditions scored 6/6. Pilot/original/
no-memory used 17/18/20 model calls, but both memory conditions cost more tokens and
API spend than no-memory. The small one-lineage ceiling result demonstrates no
accuracy improvement or learning and is not a benchmark of this stable wheel.
Specs052/053 retain all corrections, source identities, per-condition results and
costs. Historical failed task protocols were not silently rescored. Spec054 records
actual installed-user journeys on Windows/macOS/Linux, artifact/upgrade checks and
publication states. Operational passes are not production-quality benchmark evidence.
The broader Spec050 production-policy qualification, complete installed-host
conformance and efficacy roadmap remains incomplete. Qualification stays
inconclusive, unproxied host tools remain advisory, and no universal interception,
new DolphinBench result, broad superiority or automatic learning is promised.
Private/unlicensed benchmark data is not redistributed; no model weights are trained.
The separate dirty Spec051/pgvector work is excluded.
Website publication
Package publication, the private AtMem.ai source-pin PR, owner merge and live
deployment are separate states. The owner alone merges the private site PR; only
the checksum-verified artifact from validated private main may be deployed.