AgentMemory plugin

Connects Agent Zero to rohitg00/agentmemory through its local REST API.

Setup

Start AgentMemory separately:

npx -y @agentmemory/agentmemory

The plugin uses http://localhost:3111 for a host install and http://host.docker.internal:3111 inside Docker. The repository's Docker compose file already maps host.docker.internal to the host gateway. Configure the URL, secret, project-scoped behavior, and recall/capture toggles in the AgentMemory plugin settings. AGENTMEMORY_URL, AGENTMEMORY_SECRET, and AGENTMEMORY_AGENT_ID override the corresponding settings.

The plugin injects context at the start of each Agent Zero turn, records the user prompt and tool results, and closes the AgentMemory session when the turn ends. It also exposes agentmemory_search, agentmemory_save, and agentmemory_status.

Agent Zero's bundled _memory plugin is a separate memory backend. Disable one backend if both are enabled, otherwise the agent may receive duplicate recall and save the same information twice.

Memory panel (WebUI)

Adds a Memory surface to the right canvas (next to Files / Editor / Factory / Desktop):

• live server health badge (healthy / degraded / offline)
• searchable memory list (search is proxied server-side)
• recent captured sessions with observation counts
• read-only by design: the proxy whitelists health, memories, sessions, observations (GET) and search (POST); all write endpoints return 403

Panel files: extensions/webui/right-canvas-panels/agentmemory-panel.html, extensions/webui/right_canvas_register_surfaces/register-agentmemory.js, modal fallback webui/main.html + standalone webui/panel.html, backend proxy api/proxy.py. The proxy target follows the plugin url setting (or AGENTMEMORY_URL env). Requires the AgentMemory server to be running.

Auto-start server

When auto_start is enabled (default) and the configured URL points at this container (localhost/127.0.0.1/::1), the plugin keeps the AgentMemory server alive:

• At framework boot: extensions/python/startup_migration/_10_agentmemory_server.py runs ensure_server() in a daemon thread (boot never blocks).
• Self-heal: the first turn of each chat re-checks health and revives the server in the background if it died.
• The server is spawned detached (start_new_session=True) and writes to <workdir>/agentmemory.log.

Remote URLs are never spawned — auto-start only manages a container-local server. Data persists in <workdir>/data/, so memories and sessions survive restarts.

Actions tool (proactive proposals)

agentmemory_actions gives the agent first-class management of AgentMemory action items (cross-session work with priority, status, dependencies):

• create / list / update / frontier / next operations, mirroring upstream's memory_action_create/memory_action_update/memory_frontier/memory_next MCP tools
• Propose-then-confirm etiquette baked into the tool prompt: the agent proactively proposes action-worthy follow-ups at natural moments (completed work, TODOs, user-stated plans) and only creates them after user approval
• When auto_recall is on and pending actions exist, the session-start context injection surfaces the frontier (top 5) so the agent can propose continuing open work in any future chat

Relations tool

agentmemory_relations links memories so the dashboard Relations tab and memory traversal reflect how knowledge connects:

• relate — create a typed link between two memory ids (supersedes, extends, derives, contradicts, related), optional explicit confidence (otherwise computed from shared sessions/age)
• list — browse existing relations

Relations are never created automatically by the daemon (except supersedes chains from memory evolution) — the client is expected to relate the knowledge it connects. This tool is that wiring.

Memory types and procedural consolidation

agentmemory_save uses the daemon's exact type vocabulary: fact, architecture, workflow, pattern, preference, bug. Any other type silently degrades to fact upstream. pattern matters: procedural consolidation (the dashboard Procedures tab) only runs when ≥2 memories of type pattern recur (frequency ≥2) — save recurring behaviors explicitly as pattern.

LLM timeout

The daemon's default LLM timeout is 60s; consolidation prompts are large and slow providers can exceed it (intermittent Semantic consolidation failed / LLM graph extraction failed). Auto-start spawns the daemon with AGENTMEMORY_LLM_TIMEOUT_MS=120000; override via the daemon's own environment to raise or lower it.