#171 AF-280: Honcho memory layer

[Anonymous]

Originally created by: AnaniyaT

Give agents a memory backend (Honcho) that survives restarts, with per-agent observability, cost attribution, and explicit cross-agent sharing — behind a fleet-wide toggle that defaults off.

What's included

• Honcho service — new helm/honcho chart + Helmfile releases (API, deriver, pgvector DB), pinned to LiteLLM for every model call. The chart mints its own LiteLLM key in a pre-install hook, so no deploy step is manual. Releases are gated on HONCHO_ENABLED, so a memory-off deploy needs no new config.
• Runtime wiring — OpenClaw through its plugin slot, Hermes through memory.provider. Off unless HONCHO_ENABLED; running agents keep file-backed memory until their next restart.
• Memory tab (per agent) — read, correct, forget, and share what an agent learned, behind dedicated agent.memory.read / agent.memory.manage permissions. Peer filter chips with real server-side counts.
• Sharing — promote a fact into other agents; written where each runtime's recall actually looks, with provenance tracked on our side since Honcho stores none.
• Deletion — erases the agent's memory (retried on agent.deleted, since Honcho refuses a delete while a session is live), with carry-over offered in the retire dialog to copy memory elsewhere first.
• Cost — per-agent and total memory spend, attributed from Honcho's own telemetry against LiteLLM's authoritative key spend. Costs page shows a true all-in total (model + memory).
• Noise control — deriver told per-workspace to skip transient "the peer asked X" actions; OpenClaw plugin's own noisePatterns drop its memory sub-agent's scaffolding at capture.

Required repo config

New GitHub Variables/Secrets are needed only to enable memory in a deployed environment — a memory-off deploy needs none. Exact names and examples are in docs/features/af-280-honcho-memory-layer/DEPLOYMENT.md The enable flag is per-environment — STAGING_HONCHO_ENABLED (staging), HONCHO_ENABLED (prod), PUBLIC_HONCHO_ENABLED (public) — so memory can be turned on for staging first, then prod. DB user/name and model vars are shared; the DB password is per-environment.

Notes

• Draft: the memory-off deploy path renders clean, but this has not been applied to a real cluster; the org-wide memory view built earlier was removed and preserved on AF-280-org-memory-view-backup.
• Chart bumps: honcho new, agentbarn-api 0.7.13 → 0.8.0.

Testing

make check-api check-migrations check-monitoring check-ui lint-ui — all clean. make test-api (1711 passed), make test-ui (225 passed). helmfile template renders with memory off (no Honcho secrets) and on. Sharing, deletion+carry-over, cost attribution, the deriver instruction on restart, and the OpenClaw noise patterns were each verified against the running local stack.