AtMem 2.3.4

Stable release. AtMem now installs AtFlows 0.1.1 with the base package. AtFlows remains a separately started local observability service. Its optional delegated login lets one AtMem account and role govern both dashboards. AtMem also adds a bounded, read-only review of exact session-linked AtFlows trace-error leads. Neither installation nor a normal memory-health operation starts AtFlows or admits telemetry as canonical memory.

Install and upgrade

Back up or snapshot a quiescent AtMem Home before upgrading. In the Python environment used by AtMem:

python -m pip install --upgrade atmem==2.3.4
atmem status
atflows status

The base wheel pins atmem-atbot==0.1.0 and atflows==0.1.1. The older atmem[atflows] spelling remains valid. Python 3.10–3.13 and Bun 1.1+ are supported; Bun is needed only when running AtFlows. Package installation preserves existing Home, memory and evidence data, and existing AtFlows data. It does not start either dashboard, create accounts, change AtBot behavior, or enable a review job. Restart running AtMem host adapters after upgrading. Existing standalone AtFlows installations retain their local account files.

For OpenClaw, upgrade its matching bridge with:

atmem openclaw upgrade

The matching npm bridge is openclaw-memory-atmem@2.3.4. For a new connection, use atmem openclaw install.

Optional shared dashboard login

First initialize and start AtMem's dashboard. Use its own one-time Administrator password to finish AtMem account setup. Then start AtFlows with the AtMem dashboard's numeric loopback URL:

export ATFLOWS_ATMEM_AUTH_URL=http://127.0.0.1:ATMEM_PORT
atflows init

If AtFlows was initialized previously, atflows init reports that fact and exits without launching a server. In that case run atflows start with the same ATFLOWS_ATMEM_AUTH_URL setting. atflows status shows the actual dashboard and proxy ports; atmem status in 2.3.4 reports whether a server was detected but does not start one. This extra manual step is a limitation of the published 2.3.4 onboarding flow, not a required user action in the planned streamlined flow.

Replace ATMEM_PORT with the actual port printed by atmem status. Keep this variable set for every atflows start or atflows init; the mode is read at server start, not saved by init. Open both dashboards with 127.0.0.1; browser cookies are shared across ports on that host. Sign in at AtMem, then open AtFlows. AtMem owns accounts, passwords, roles, session revocation and user management in this mode. AtFlows hides its local user-management controls; signing out of AtFlows ends the shared session. The authority must be a same-host numeric-loopback HTTP origin. When AtMem is unavailable or a session is revoked, AtFlows protected requests are denied. Remove the environment setting and restart AtFlows to use its existing standalone accounts again. There is no account migration or automatic mode switch.

Optional evidence review

An authorized AtMem Investigator, Evidence Collector or Administrator may review an exact run/session pair and time window:

atmem atflows review RUN_ID --session-id SESSION_ID \
  --since-ms START_MS --until-ms END_MS \
  --base-url http://127.0.0.1:1337

The command requests an AtMem evidence token (or reads ATMEM_EVIDENCE_TOKEN). With standalone AtFlows it asks for the AtFlows Administrator password (or reads ATFLOWS_ADMIN_PASSWORD). With delegated AtFlows it asks for an AtMem username and password (or reads ATMEM_USERNAME and ATMEM_PASSWORD), creates a short-lived local session for the query, then revokes that session. --human prints a concise view; the default is versioned JSON. It reads at most 100 trace rows from a numeric-loopback endpoint in a window of at most seven days. Only exact session links in authorized AtMem run evidence produce leads. Raw traces, prompts and error messages are not copied into the report. Missing links and missing coverage remain visible as gaps.

Compatibility and limits

AtMem retrieval, encrypted evidence, Home format and AtBot behavior remain compatible with 2.3.4b1. The previous beta's local AtFlows handoff gains an AtMem-owned login option; standalone AtFlows remains supported. AtFlows telemetry is observational: a trace error is a lead, not proof that injected memory caused an outcome. The Memory Health queue, background review, automatic proposals and causal experiments remain future work. AtFlows does not automatically ingest AtMem traces, and the two products do not share a database. External side effects still require independent receipts.