AtMem 2.3.3 — more efficient retrieval, clearer investigation

AtMem 2.3.3 builds on the encrypted, portable Agent Black Box introduced in 2.3.2. This stable release reduces avoidable retrieval work, tightens a personal fact support check, and makes session evidence easier to navigate. Memory authority, scoped access and exact delivery remain the foundation.

Three improvements that matter

1. Less unnecessary work when finding memory

Common deterministic query expansions now run locally rather than requiring a companion request. Exact semantic search benefits from batched record validation, bounded query-embedding reuse and optional NumPy scoring. Compatible unchanged vectors can be reused during a staged index rebuild, avoiding redundant embedding calls while preserving model identity and canonical-state checks.

These are targeted improvements, not a change to approximate nearest-neighbor search. Gains depend on corpus size, embedding backend and workload; this release does not promise a universal latency multiplier.

2. More deliberate context selection

An explicit question about the user's own age cannot establish direct support solely from a relative's age. Unicode-aware token handling and retrieval regression coverage strengthen the existing support checks. Authorization and final canonical revalidation continue to govern which memory reaches an agent.

Developers can evaluate a new core-rrf-v1 candidate strategy that combines independent lexical, fact-key, semantic and bounded graph nominations. It records channel contributions and graph support paths, filters by scope before ranking, and revalidates supporting records. It is opt-in, not the default for adapters.

3. A session archive that fits the available space

Archive filters and session cards now respond to the width of their panel, including when the evidence inspector is open. Long identifiers no longer force oversized cards, and actions remain within the pane. Browser layout checks cover wide desktop through narrow mobile widths.

Install or upgrade

New installation:

python -m pip install atmem==2.3.3
atmem atbot setup
atmem init
atmem dashboard

Upgrade an existing installation:

python -m pip install --upgrade atmem==2.3.3
atmem openclaw upgrade       # if OpenClaw is connected
atmem dashboard daemon restart
atmem control verify
atmem home status

Restart other long-running AtMem clients after upgrading. For optional framework integrations, install the matching extra:

python -m pip install --upgrade 'atmem[pydantic-ai]==2.3.3'
python -m pip install --upgrade 'atmem[langgraph]==2.3.3'

The matching npm bridge is openclaw-memory-atmem@2.3.3. AtBot remains the unchanged, independently versioned stable dependency atmem-atbot==0.1.0. None of these package versions is a beta.

Defaults, migration and optional features

• Existing retrieval strategy and adapter defaults are unchanged. The new fusion path requires an explicit RecallRequest(retrieval_strategy="core-rrf-v1", ...). Its graph is associative, not causal; ranking does not itself authorize injection.
• Optional NumPy scoring is available with python -m pip install --upgrade 'atmem[semantic-accelerated]==2.3.3'. Instance-local vector matrix reuse additionally requires SemanticIndex(..., cache_vectors=True) and remains off by default. It is bounded to one entry and 32 MiB of retained matrix/blob storage, not total process memory. Short-lived native index instances do not gain cross-turn reuse.
• There is no new persisted schema migration for these retrieval changes. Existing Home locations, accounts, encryption, capture settings and delegated registrations are retained. HMAC request authentication and signed delegated results are unchanged; native ranking does not rewrite provider-owned context.
• Before upgrading, make a verified Home snapshot or stop all writers before copying the Home. Keep it until the upgrade is verified. To revert the package, install atmem==2.3.2, run atmem openclaw upgrade if connected, and restart the dashboard/clients. Return experimental callers to the legacy request profile. Restore a snapshot only deliberately: it excludes work recorded afterward.

Compatibility and validation

Python 3.10–3.13 remain supported. The release workflow checks Python and AtBot, optional framework/provider integrations, the OpenClaw bridge, distribution metadata, installed wheels and persisted upgrades, including from 2.3.2. OpenClaw's locked-host conformance profile uses 2026.8.1; compatibility with every future host version is not implied. The base install does not require NumPy, a hosted model, or another memory engine.

The release retains encrypted evidence, local role enforcement, standalone Home recovery, and native/delegated context delivery. Full-fidelity multimodal host capture remains verified for the supported OpenClaw profile; framework delivery support does not imply identical multimodal capture coverage.

Opt-in options and evidence scope

Fusion, scoped graph nominations and matrix reuse ship as developer evaluation options. Upgrading does not enable them. Calibration reports, including measured limits, are in the benchmark guide.

AtMem preserves what supported agent boundaries supply. An event records a host-observed action; an external purchase or message carries its own receipt. Content that was never captured stays uncaptured — fidelity is decided at capture time. These evidence boundaries are unchanged from 2.3.2.