agentmemory-architecture

어떤 스킬인가요?

How agentmemory is built, the iii engine primitives it runs on, its storage model, ports, and the viewer. Use when reasoning about how memory is stored or retrieved end to end, when extending the system, or when answering how agentmemory works under the hood.

설치 방법

프로젝트 폴더에서 아래 명령어를 실행하고, 설치할 에이전트를 선택하세요.

설치 명령을 직접 실행해야 적용됩니다. 지원 에이전트와 필요한 권한·라이선스는 제작자의 안내를 확인하세요.

스킬 원문 파일

---
name: agentmemory-architecture
description: How agentmemory is built, the iii engine primitives it runs on, its storage model, ports, and the viewer. Use when reasoning about how memory is stored or retrieved end to end, when extending the system, or when answering how agentmemory works under the hood.
user-invocable: false
---

agentmemory is a memory server for coding agents. It runs locally, captures observations, indexes them for hybrid retrieval, and serves them back over REST and MCP. Worker Compose runs the pinned iii engine with AgentMemory and the registry `state` and `http` workers. The state worker uses a file-backed KV store by default, with Redis available by configuration; it is not SQLite.

## iii primitives

Everything is a function, a trigger, or worker state on the iii engine. There is no separate plugin system; the worker registers functions (`mem::*`) and HTTP triggers (`api::*`) and the engine routes calls. agentmemory does not bypass iii; new capability is a new function plus a trigger.

## Retrieval model

Recall combines BM25 keyword search with configured vector similarity and graph expansion. A clean installation has no embedding provider and uses BM25 without an API key. Set `EMBEDDING_PROVIDER=local` to enable on-device embeddings, which download a model once. A remote embedding provider runs only when `EMBEDDING_PROVIDER` names it. LLM summaries, consolidation and reflection run only when `AGENTMEMORY_LLM_PROVIDER` selects a provider or `AGENTMEMORY_ALLOW_AGENT_SDK=true` turns on the Claude Agent SDK fallback; a key in the environment alone does not turn them on. Per-observation LLM compression also needs `AGENTMEMORY_AUTO_COMPRESS=true`.

## Data boundary

Captured prompts, tool inputs and outputs, and memories stay in the local store. They leave the machine only through options the user turns on: a selected LLM provider, the Agent SDK fallback (`AGENTMEMORY_ALLOW_AGENT_SDK=true`), a remote embedding provider, a Redis state backend, or mesh sync to peer servers the user registers.

## Storage and lifecycle

Memories carry content, concepts, files, importance, and timestamps, grouped into sessions and optionally linked to commits. A lifecycle of capture, compress, consolidate, and forget keeps the store useful over time rather than letting it grow unbounded.

## Ports

REST defaults to port `3111`, the viewer to `3113`, and the internal engine WebSocket to `49134`. Use `--port <N>` or `--instance <N>` to select another local runtime.

## Viewer

A real-time web viewer at `http://localhost:3113` shows memory building as sessions run, with authenticated SSE updates over the REST port. Useful for demos and for confirming capture is working.

## See also

- agentmemory-mcp-tools and agentmemory-rest-api for the surfaces.
- agentmemory-hooks for automatic capture.
- agentmemory-config for ports and feature flags.

<!-- AUTOGEN:troubleshooting START - generated by scripts/skills/generate.ts, do not edit by hand -->
# Troubleshooting agentmemory skills

Recovery steps for separately installed skills. This document is generated into
each skill directory so native skill readers can open it without parent traversal.

## "MCP tool not available"

Missing tools can mean a stopped MCP process, a disconnected daemon, or an
explicitly reduced tool surface. Walk these in order:

1. Identify the host. In Claude Code, inspect `/plugin list` and `/mcp`.
   In Hermes, inspect the configured memory provider and `mcp_servers` separately.
   In OpenClaw, validate the host configuration and inspect its MCP servers and
   native plugins separately. In Pi or OMP, inspect `/agentmemory-status` and the
   MCP inventory separately; native save/search does not prove the broader tools
   are connected. Do not run another host's setup commands.
2. Restart or reload the affected host connection after configuration changes.
3. Inspect the connected tool inventory. Hermes may namespace MCP tools; invoke
   the actual advertised names and schemas. A native memory provider is not
   proof that every MCP tool or shared skill is available.
4. Check `agentmemory status` and the daemon's `AGENTMEMORY_TOOLS` setting.
   Codex's bundled bridge lists the daemon's enabled tools. The standalone shim
   used by other hosts can fall back to seven basic tools with a separate store.
   Do not claim full daemon functionality based on those seven tools.
5. If `agentmemory connect codex` and the Codex plugin are both enabled, choose
   one MCP connection. Check `/hooks` before adding global hook fallbacks.

Hermes, OpenClaw, Pi and OMP connector installations bundle skills; other hosts
may install them separately. Check the host's skill inventory and open this file
from the installed skill directory. Test automatic capture with a real turn and
a recorded observation; tool connectivity alone does not establish lifecycle
capture.

An outage or timeout is not an empty search result. Report the failure. A write
may have committed before the response was lost; inspect state before retrying.

## REST fallback

When the MCP tools stay unavailable but the daemon is running, call the REST
API directly:

1. Set `AGENTMEMORY_URL` to the daemon base URL (default `http://localhost:3111`).
2. The REST API requires a bearer credential by default. Use the host's explicit
   `AGENTMEMORY_SECRET`; for loopback URLs only, fall back to the secret in
   `~/.agentmemory/.env` and then `~/.agentmemory/secret`. Load the value locally
   without displaying it. Never send a local-file secret to a remote URL.
   If `AGENTMEMORY_CREDENTIAL_MODE=explicit`, require the explicit host credential
   and do not read local credential files.
   Authenticated REST fallback requests require HTTPS for non-loopback
   URLs; HTTP is allowed for loopback URLs (`localhost`, `127.0.0.0/8`, or `[::1]`).
   Reject redirects on authenticated requests. Do not print or save the secret.

Endpoint map by skill:

| Skill           | REST call                                                        |
| --------------- | --------------------------------------------------------------- |
| remember        | `POST /agentmemory/remember`                                     |
| recall          | `POST /agentmemory/smart-search`                                 |
| recap           | `GET /agentmemory/sessions` + `POST /agentmemory/smart-search`   |
| handoff         | `GET /agentmemory/sessions` + `POST /agentmemory/smart-search`   |
| session-history | `GET /agentmemory/sessions`                                      |
| commit-context  | `GET /agentmemory/session/by-commit?sha=<sha>`                   |
| commit-history  | `GET /agentmemory/commits` (URL-encode every query param)       |

The MCP host reads the plugin MCP configuration. Restart the daemon after
changing its configuration, then restart or reconnect the host MCP process after
changing its URL or explicit secret. The local bridge reads the credential files
for each request; it does not inherit other daemon `.env` settings.
<!-- AUTOGEN:troubleshooting END -->