AtMem 2.3.15b1 — Governed RAG and verified feedback beta

This beta adds opt-in, scoped document retrieval and feedback investigation while preserving ordinary local memory/context defaults. It is a development candidate, not a stable promotion, production qualification or a new DolphinBench result. AtMem 2.3.14 remains the stable line.

Install or upgrade

Use Python 3.10–3.13 in a separate virtual environment for beta testing:

python -m pip install --upgrade 'atmem[code]==2.3.15b1'
python -m pip check
atmem --version
atmem status

The minimal package is python -m pip install --upgrade atmem==2.3.15b1. The code extra supplies SQLCipher and recipient cryptography for the fuller paths. No model API or paid evaluation is started by installation.

Exact compatibility: atmem==2.3.15b1, atmem-atbot==0.1.1 (unchanged), atflows==0.1.8 (unchanged), openclaw-memory-atmem@2.3.15-beta.1. The bridge publishes under npm beta, not latest. To update a beta bridge, use atmem openclaw upgrade, then atmem openclaw status.

Back up the complete AtMem Home, canonical database, encryption policy, keys, identity/accounts and feedback checkpoints before opting in. Finish active work, stop services you own, install the beta and restart clients. Test synthetic data first. Preserve the backup for rollback; do not point an older writer at a Home that has opted into the new feedback container format.

What is implemented

• Owner-registered RAG providers and a local encrypted-vault/RAM-FTS document adapter, returning original cited passages, source revisions and requested jurisdiction/effective-date filters. No bundled law corpus or legal accuracy guarantee is included.
• A separate authenticated local stdio MCP server and CLI for scoped retrieval, read/report/propose operations and registered guarded proxy handlers. It is generic MCP integration; Hermes is optional. Legacy MCP defaults remain intact.
• Per-action read/write outcomes, independent backend-effect readback, preserved contradictions and bounded operational failure classifications. Missing facts, retrieval misses, delivered-but-not-applied evidence and wrong argument values need their own actual proof; a trace or model label cannot establish them.
• Reviewed typed rules and retrieval profiles, shadow reporting, exact single-use approvals, explicit revocation, uncertain-write holds and recipient-encrypted, consented example export. Dashboard, CLI and MCP share authority state.
• Frozen diagnostic comparisons pin verifier generations, original input and trial provenance, retaining exact outcome/verification identities. Missing, changed or disputed evidence remains unknown.

Start with the governed RAG/MCP guide and verified-feedback setup and limits. Owner configuration, independent readback code and explicit agent grants are required; installing this package does not automatically capture conversations, connect an external database or authorize writes.

Fixes and recovery

Windows migration flushes preserve binary database bytes, including a trailing 0x1A. Feedback checkpoint paths work with classic MAX_PATH settings without a machine-wide registry change. Installed-package checks identify the actual installed distribution rather than accidentally importing a source checkout.

Canonical backup/restore uses the active SQLite/SQLCipher driver and household key; encrypted snapshots remain ciphertext. Supported direct restores publish private content-free pending/completed witnesses. Attached feedback stays blocked until a completed restore is explicitly reconciled by its owner. Reconciliation disables old grants, approvals and consents and quarantines restored policy; it does not authorize activation. Interrupted restores require explicit retry from a known-good snapshot. Native recall remains available.

Beta boundaries

Production qualification still issues inconclusive diagnostic receipts. Independent power/corpus/resource acceptance, complete scoped installed-host conformance, the final in-flight/deletion acceptance matrix and a separately approved efficacy evaluation remain unfinished. No public tool bypasses these gates. The pinned Hermes integration is a development candidate and has not replaced the owner's installed runtime. External/unproxied host tools remain advisory; MCP connectivity does not control unrelated tools.

This beta does not train a model, claim general causality or improve the frozen DolphinBench score by implication. Synthetic tests exercise real storage and local effects, but do not establish production quality, scale or superiority. A complete rollback of Home, keys and checkpoints lacks an external monotonic witness; arbitrary filesystem copies need operator reconciliation. Withdrawn export consent cannot recall uncontrolled copies or erase trained weights.

Enabling feedback opts its existing evidence container into format 4. Old readers must refuse that container instead of modifying it. Ordinary unused containers and native memory retain their existing behavior. To roll back, restore the complete pre-opt-in backup into a separate Home and follow the supported adoption/reconciliation procedure; there is no in-place downgrade.

Website documentation publication is separate: the source-pin update requires an owner-reviewed website PR, followed by an explicitly authorized deployment. A package publication alone does not update the live website.