feat: agent memory & work log (Phase D)

[michaelzwang13]

Phase D — Agent memory & work log

Closes the memory loop and persists the audit trail. Phase B left a deny-only log line in policy.require_action; Phase D promotes that stub to a persisted agent_action_log row covering both allow and deny. Adds a per-agent key/value memory store the agent writes via a new update-memory skill and reads back as injected role_context on every dispatch — knowledge now survives container restarts. Pulls Phase C's reviewed_prs dedup index forward so the autonomous PR-review loop can build on a stable base.

Schema (migration 004 + schema.sql)

• agent_memory (agent_id, key, value, updated_at) — unique(agent_id, key)
• agent_action_log (agent_id, action, outcome, metadata jsonb, created_at) with a CHECK constraint pinning outcome to allowed | denied
• reviewed_prs (agent_id, owner, repo, pr_number, reviewed_at) with unique(agent_id, owner, repo, pr_number)
• RLS closed-by-default policies mirrored across both files

Models

• AgentMemoryModel.get / upsert / list_by_agent / delete
• ActionLogModel.record / list_by_agent
• ReviewedPRModel.exists / record / list_by_agent

Policy

• require_action persists allow + deny rows. Audit writes are best-effort — a DB hiccup logs locally and lets the request proceed, trading audit completeness for availability; the policy decision itself never depends on the DB.

Dispatcher

• dispatch_task now loads the calling agent's memory and injects it into role_context (best-effort; a memory-load failure does not block dispatch).

Gateway

• GET / POST /gateway/memory — agent-authed, policy-gated via agent.memory.{read,write}. Per-agent only — memory is scoped by agent_id.
• POST /github/review now inserts a reviewed_prs row on success, guarded by exists() so re-reviews don't double-write.

Skill + template

• New update-memory skill — agent POSTs key/value pairs via the agent bearer token. Server-side persistence is what lets preferences survive container restart (SOUL.md cannot).
• code-review-engineer.yaml: skills += update-memory; allowed_actions += agent.memory.read, agent.memory.write.

Tests

125/125 pass. 11 new:

• policy — audit allow + deny + audit-failure swallowed
• gateway — memory write/read/auth/policy/empty-key; reviewed_prs records on success + skips on dup
• dispatcher — memory injection + best-effort failure

Live smoke (12/12 green against the running backend)

Write/read round-trip; upsert idempotent on key; 401 on bad/missing token; 403 when a secretary role hits /memory; agent_action_log carries the allow + deny rows with role metadata.

Operator note

Migration 004 needs to be re-run in the Supabase SQL Editor to pick up the three new tables (idempotent — create table if not exists makes re-runs safe). Agent image was rebuilt locally to include the new skill — docker build -t openclaw/agent:latest backend/agent-runtime/.

Out of scope, follow-up

Memory compaction (LLM reflection / LRU / clustering) is tracked in #23 — the forward-compatible defaults this PR sets (updated_at per row, all-keys injection at dispatch) keep that path open without schema changes.

Closes #8.

🤖 Generated with Claude Code

[michaelzwang13]

@claude review

[claude]

Missing doc updates — CLAUDE.md compliance

CLAUDE.md L87 requires:

When behavior or setup changes, update the relevant md (README.md, LOCAL_SETUP.md, ROADMAP.md, this file) in the same change.

This PR introduces three categories of changes that cross that threshold but no doc files are updated:

1. Setup change — LOCAL_SETUP.md should note that migration 004 must be re-run in Supabase to create agent_memory, agent_action_log, and reviewed_prs. The operator note lives only in the PR description right now and won't be discoverable after merge.
2. Architecture change — CLAUDE.md's architecture section lists the backend routers/services. The two new GET /gateway/memory and POST /gateway/memory endpoints and the three new model files (AgentMemoryModel, ActionLogModel, ReviewedPRModel) are worth a line there.
3. Behavioral change — agents now carry persistent memory across container restarts via the update-memory skill; this is non-trivial new platform behavior worth a mention in README.md.