AtMem 2.3.11b1 — experimental action-routing candidate

This prerelease packages experimental action-routing and source-linked standing facts alongside the existing AtMem memory and context-provider paths. These features are not activated by ordinary installation. This candidate is intended for future qualification; it does not imply a new full DolphinBench result. The stable release remains AtMem 2.3.10.

Install and compatibility

Use Python 3.10–3.13 in an isolated environment:

python -m pip install --upgrade --pre atmem==2.3.11b1
atmem --version
atmem status --json

AtMem pins atmem-atbot==0.1.1 and the published prerelease atflows==0.1.5b1. The managed OpenClaw bridge is openclaw-memory-atmem@2.3.11-beta.1; after upgrading Python, use atmem openclaw upgrade for an existing installation or atmem openclaw install for a new one. Do not install the bridge separately. Optional encrypted coding memory uses python -m pip install --upgrade --pre "atmem[code]==2.3.11b1". The optional Copilot Code host also accepts the locally tested CLI 1.0.103 on macOS; Windows remains qualified for 1.0.98 only. Other versions fail closed.

Back up your AtMem Home, database, keys and evidence before upgrading; restart long-running clients. The existing data migration and rollback guidance in the 2.3.10 release note applies. Host routing and standing-fact experiments require separate configuration and are not automatic.

Research limits

The action guard is a separately staged Hermes plugin that can propose routes, hold a first write for context inspection, and report unconfirmed operations. It is not a guarantee of task success, complete policy enforcement, safe replay, or support for arbitrary application tools. The benchmark runner, its source data and its qualification receipts are maintained separately from this package.

The experimental standing-fact extractor checks whether a source quotation occurs in a supplied history. That does not verify that a generated statement is supported by or consistent with the quotation. Its derived stores and MCP server are not the normal governed AtMem memory store or an authorization boundary. Do not rely on generated facts without reviewing their original source. The research stores and their use can involve sensitive data and external model calls when explicitly configured; assess data handling before use.

The earlier 17/20 result was a post hoc diagnostic on selected tasks with a different frozen artifact and condition; it is not a score for this package or proof that memory learned. The original 390/600 result remains unchanged. Before any new 600-task campaign, fresh identity-bound source, product, profile, memory and environment qualification must pass at the actual launch boundary.

Website source-pin refresh, owner review and deployment are separate from package publication; live website documentation is pending independent verification.