agentmem

Vectorless memory for AI agents.

Agents forget. Worse: when memory lives inside a model provider or a single harness, your conventions, decisions, and project knowledge leave with the product.

agentmem is a pluggable memory layer you own. Plain markdown on your machine, one memory binary, wired into any agent that can run shell or MCP — so the memory stays yours when you switch tools.

Free for personal and commercial use. Source is private. This repository holds the documentation site, issue tracker, and release binaries.

🌐 อ่านภาษาไทย ที่นี่

Install

curl -fsSL https://agentmem.thamtech.co/install | sh

Or download the latest binary from GitHub Releases.

Wire a harness

Official harnesses:

memory install claude-code
memory install opencode

MCP (also registers the MCP server; the normal guidance remains installed):

memory install --mcp claude-code
memory install --mcp opencode

Other / skill-capable harnesses (portable skill only — no harness-specific integration):

memory install skill
# optional custom skills root:
memory install skill --dir /path/to/skills

Default skill path: ~/.agents/skills/memory/SKILL.md (shared with OpenCode’s skill root).

Check status: memory targets.

Why

• Yours — markdown files on disk, not locked in a provider account
• Pluggable — skill or MCP into the harness you already use
• Portable — switch agents; keep the same memory folder
• Vectorless — no embeddings, no daemon, no cloud required

Docs site

A small Go server embeds the entire site/ tree (//go:embed) into one portable binary — templates, CSS, JS, assets, language packs, and Markdown.

i18n

All UI copy lives in YAML under site/lang/. Root keys are page names (common, index, install, commands, integrations, how_it_works).

# site/lang/en.yaml
common:
  nav:
    install: Install
index:
  title: "agentmem — vectorless memory for AI agents"
  tagline: Vectorless memory for AI agents

Templates call {{.T "key"}} (page-local, then common) or absolute paths {{.T "common.nav.install"}}.

Language selection (first match wins):

1. ?lang=en
2. Cookie lang=en
3. Accept-Language
4. Default en

Add a locale by copying site/lang/en.yaml → site/lang/<code>.yaml and translating values.

Route
/	Landing
/docs/install	Binary + harness wiring
/docs/commands	CLI reference
/docs/integrations	Claude Code, OpenCode, other harnesses
/docs/how-it-works	Storage, index, local UI
/*.md / /docs/*.md	Same pages as Markdown (Content-Type: text/markdown)

Makefile

make build                 # → ./agentmem-site
make install               # → ~/.local/bin/agentmem-site
make run                   # go run . -addr :5555  (ADDR=:8080 to override)
make test
make fmt
VERSION=0.1.0 make build   # stamp version into binary
./agentmem-site -version

Manual build / run

# preview (rebuilds embed from site/)
go run . -addr :5555

# portable binary (no site/ dir needed at runtime)
make build
./agentmem-site -addr :5555
# ship only agentmem-site to the server

Layout: site/templates/ (layout, header, footer + page bodies).
Static: site/css, site/js, site/assets. Markdown: site/index.md, site/docs/*.md.
Content changes require a rebuild so they are re-embedded.

Core commands

memory list [prefix]          # show the memory index, optionally scoped
memory show <name>            # read one memory
memory search <query>         # grep across memories
memory add <text>             # append a quick note
memory add --file <name> --summary "[tag] desc" <text>
memory edit <name>            # open in $EDITOR
memory rm <name>              # delete a memory
memory reindex                # rebuild MEMORY.md
memory serve                  # MCP server over stdio
memory ui                     # local loopback web UI
memory install <harness>      # Claude Code: skill + CLAUDE.md; OpenCode: skill + plugin
memory install --mcp <harness>
memory install skill [--dir <dir>]
memory uninstall …
memory targets                # harness + skill install status

How it works

Everything lives in ~/.agentmem/ by default (override with AGENTMEM_DIR). Each memory is a markdown file. MEMORY.md is a generated index from each file’s summary line. Files are the source of truth.

Supported harnesses

Target	What you get
Claude Code (official)	Skill + managed CLAUDE.md guidance · optional MCP
OpenCode (official)	Skill + plugin · optional MCP
skill (portable)	Skill only under ~/.agents/skills — no harness-specific integration

Issues & releases

• Bugs and features: GitHub Issues
• Binaries: GitHub Releases

License

See LICENSE. Free to use; source is not open.