AtMem 2.3.9 — Shared project memory for coding agents

AtMem Code is an optional flavour for source-backed project navigation and persistent coding experience. Share authorized project memory between agents, inspect earlier decisions and command outcomes, explore code relationships and move encrypted project memory to another clone.

Install or upgrade

python -m pip install --upgrade atmem==2.3.9

To add Code:

python -m pip install --upgrade "atmem[code]==2.3.9"

Follow the five-step Code guide. Standard installation keeps Code disabled. No Laya model, learned skimmer, Torch or GPU is required. Richer multi-language extraction uses an explicitly built and approved local Docker worker; the Python syntax index runs locally without Docker.

AtMem pins stable atflows==0.1.4 and the existing stable atmem-atbot==0.1.0. AtBot does not need separate installation. Installing packages does not start servers or collect telemetry. Use atmem init and atmem status for dashboard startup and actual URLs. AtFlows remains a separate telemetry store.

OpenClaw users update the matching bridge and restart their gateway:

openclaw plugins install openclaw-memory-atmem@2.3.9

What Code adds

• Shared project/profile memory across authorized coding agents.
• Source-backed symbols, relationships, architecture and potential change impact.
• Searchable experience with recorded intent, source anchors and observed outcome.
• Index-first lookup: request selected evidence instead of automatic source dumps.
• CLI/dashboard search, inspection, correction, scoped deletion and pause controls.
• Signed encrypted project export/import and explicit destination activation.

Automatic hooks and read-only MCP are separate. Host activation verifies the installed boundary; account policy can prevent MCP access. An MCP connection does not force tool use or replace all native exploration. See the guide for exact setup and the structure reference for optional tools.

Compatibility and migration

Python 3.10–3.13 remain supported. Existing deterministic memory, integrity and scope checks remain in force. There is no automatic conversion of ordinary memory into Code memory and no new canonical-memory schema migration. Back up existing households and stop writers before upgrading, then restart clients.

Code requires encrypted storage. Plaintext households need explicit migration. Existing Code settings retain their presentation mode; fresh projects use a small experience index. Disable stops automatic capture/recall without deleting data. Rich indexes need reindexing after source or worker changes. New bundle formats require a compatible reader; keep the previous backup before downgrading.

Evidence

Browse the AtMem/Cognee/controlled-native comparison, including all 18 scored attempts, tokens and timing. The report explains the separate Click conditions and one-task SWE-bench qualification. This is diagnostic evidence, not a full benchmark leaderboard result.

Release checks are tracked in the release contract and published-package workflow. Package publication, website owner review and live website deployment are verified separately; a source version change alone is not publication.