agentmemory-architecture Skill (agentmemory) for Claude: What It Does and How to Install It

Explains to Claude how the agentmemory memory server is built, stores memories and finds them again.

Installs: 11K; on skills.sh
Stars: 30.1K; on GitHub
Files: 1; besides SKILL.md
Updated: Oct 2026; last repo commit

This is a reference skill for agentmemory, a memory server for coding agents. It says the server runs locally, captures observations, indexes them for hybrid retrieval and serves them back over REST and MCP. It covers the iii engine primitives, the retrieval model, the storage lifecycle, the default ports and the real-time viewer.

Claude loads it on its own, because the file sets user-invocable: false. That happens when it reasons about how memory is stored or retrieved, extends the system or explains how agentmemory works. It describes the system and runs nothing itself. The memory server it describes must be running for any of it to matter.

Good for

Developers extending or debugging agentmemory
Anyone asking how agentmemory stores and retrieves memory under the hood
People checking ports, the viewer and what runs locally

Install agentmemory-architecture

Claude Code

Copy this prompt.

Install the agentmemory-architecture skill in Claude Code. Run this command:

npx -y @agentmemory/agentmemory@latest

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

Requires Node.js 20 or newer.

This skill describes how agentmemory works, so it is only useful with the memory server running. The plugin route is the one the README recommends for Claude Code. The skills-only route installs all 17 skills but does not start the server. Pick one route, not both.

From the project README. Source.

What's in the skill

When it loads

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.

Sections in SKILL.md

iii primitives
Retrieval model
Storage and lifecycle
Ports
Viewer
See also

Key facts

Author: Rohit Ghumare (rohitg00), per plugin.json
License: Apache-2.0 (repo and plugin.json)
Installs on skills.sh: 11,037 (Nov 2, 2026)
GitHub stars: 30,069 for the whole rohitg00/agentmemory repo (Nov 2, 2026)
Last repo commit: Oct 29, 2026
What it includes: 2 files including SKILL.md: TROUBLESHOOTING.md (no scripts)
Part of: The agentmemory plugin (17 skills, 12 hooks, 54 MCP tools per plugin.json)
Needs: Nothing to run by itself. It describes the agentmemory server, which you start separately

What it includes

Everything is a function, a trigger or worker state
The worker registers mem::* functions and api::* HTTP triggers on the iii engine. New capability means a new function plus a trigger.
Try: “How does agentmemory store and retrieve a memory end to end?”
Read · iii primitives
BM25 keyword search plus vectors and graph expansion
Without an embedding provider it uses BM25 and needs no API key. EMBEDDING_PROVIDER=local turns on on-device embeddings, which may download a model.
Read · Retrieval model
Capture, compress, consolidate and forget
Memories carry content, concepts, files, importance and timestamps, grouped into sessions and optionally linked to commits. The state worker uses a file-backed store by default, with Redis by configuration.
Read · Storage and lifecycle
REST 3111, viewer 3113, engine WebSocket 49134
Use --port or --instance to select another local runtime.
Read · Ports
Real-time web viewer at http://localhost:3113
Shows memory building as sessions run, with authenticated SSE updates over the REST port.
Read · Viewer
Points to the MCP tools, REST API, hooks and config skills
The sibling skills are agentmemory-mcp-tools, agentmemory-rest-api, agentmemory-hooks and agentmemory-config.
Read · See also

From the official docs

Permissions

Medium risk

The skill only explains the system, but the plugin around it records your sessions through hooks and starts a local server. Review what the viewer shows is stored.

What it can do

Read your data
Create and change things
Delete things
Read-only mode

Watch out for

The plugin records your sessions through hooks
The README says the Claude Code hooks capture user prompts (privacy-filtered), tool use and results, errors and the assistant's reply for each turn. Review what is stored in the viewer at localhost:3113 before you rely on it. github.com
Setup runs npm code and starts a local server
The README's setup command is npx -y @agentmemory/agentmemory@latest, which pulls the latest package and starts the memory server. The plugin's .mcp.json also starts a node bridge script from the plugin folder. github.com
Local embeddings download a model
SKILL.md says EMBEDDING_PROVIDER=local enables on-device embeddings, which may download a model. LLM compression and summaries need a provider you configure. github.com

Security advisories

No published security advisories as of Nov 2, 2026.

Technical details

Turning on read-only	No read-only switch is documented for these skills.
Talks to	The server and viewer bind to loopback by default, per the README. A model download if you set EMBEDDING_PROVIDER=local. A provider you configure for LLM compression or summaries. The hook scripts call to AGENTMEMORY_URL, which defaults to http://localhost:3111; if you point it at a remote server, hook data goes there
Sign-in and keys	A local server makes a random secret at ~/.agentmemory/secret (mode 0600, per the README) and bundled clients read it. A remote server needs AGENTMEMORY_SECRET in your environment.
Scopes	The Claude Code plugin registers hooks for 12 events (such as every prompt, every tool call and session end), each running a node script from the plugin The plugin ships a .mcp.json that starts a node bridge (plugin-bridge.mjs) to the memory server Setup runs npx -y @agentmemory/agentmemory@latest, which starts a server on localhost:3111 with a viewer on localhost:3113 This skill only describes the system; its SKILL.md defines no commands to run In the plugin source, session-start.mjs and pre-tool-use.mjs read AGENTMEMORY_URL and the bearer secret (AGENTMEMORY_SECRET, an .env file or ~/.agentmemory/secret) and call the memory server; session-start.mjs and stop.mjs also run git rev-parse

Questions

What does the agentmemory-architecture skill do?

It gives Claude a short description of how agentmemory works. Claude uses it to answer questions about storage, retrieval, ports and the viewer, and to extend the system the way it was designed.

How do I install it?

Use the install box at the top of this page. It is part of the agentmemory plugin, which needs the memory server running. The skill is a description, so it does nothing without the server.

How does agentmemory find memories?

Recall combines BM25 keyword search with vector similarity, when an embedding provider is configured, and graph expansion. A clean install without an embedding provider uses BM25 only, with no API key.

Is the store SQLite?

No. SKILL.md says the state worker uses a file-backed key-value store by default, with Redis available by configuration, and that it is not SQLite.

Which ports does it use?

REST on 3111, the viewer on 3113 and the internal engine WebSocket on 49134. Change them with --port or --instance.

Is it safe?

The skill itself is text and bundles no scripts. The plugin that carries it is larger: it adds hooks that record your sessions and starts a local server through npm. See the safety section on this page.

Information

updated Nov 2, 2026

Repository: rohitg00/agentmemory
Folder: plugin/skills/agentmemory-architecture

rohitg00/agentmemory skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 7

agentmemory README · github.com
skills CLI commands · github.com
agentmemory-architecture SKILL.md · github.com
plugin.json · github.com
plugin hooks.json · github.com
plugin .mcp.json · github.com
pre-tool-use.mjs · github.com