Agenticmemory

agmry

[npm version] [MCP] [Remote MCP] [smithery badge]

Persistent memory for AI agents — conversation history, key-value context, and semantic search across sessions. As easy as git.

git remembers your code. agmry remembers everything else.

Your agent writes great code all day — then forgets every decision the moment its context window resets. So you become the memory layer: re-explaining the project, the preferences, what broke last time. Agentic Memory is the memory your agent runs itself: one install, and it stores, recalls, and searches its own state across sessions, machines, and even other agents. It can even sign itself up — one command returns a working API key, no browser, no human.

Works with: Claude Code · Cursor · Cline · Windsurf · Aider · Codex · any MCP client

[agmry demo — install to full recall in 60 seconds]

Install to full recall in 60 seconds — real terminal session, no mockups. Watch all the demos.

Install

npm install -g agmry

Quick Start

# Agent signs itself up — working key returned instantly, no card, no browseragmry signup my-project --local
# Store memoryagmry store my-project "User prefers TypeScript strict mode and pnpm"
# New session? Recall everythingagmry recall my-project
# Half-remember something? Search itagmry search my-project "how do we deploy"
# Full referenceagmry --help

Agents sign themselves up

No browser. No OAuth. No waiting for a human.

$ agmry signup my-project✓ Project "my-project" created  API key: amk_************************  (saved to .agmry/config.json)  Live:    full access for 48h — no card, $0 today

The key works immediately — every endpoint, every tool, full access for 48 hours. Within that window, your human completes a $0 card-auth that starts the free 7-day trial on the same key. Miss the window? Nothing is deleted — the key pauses and revives the moment the card-auth completes. Add --email [email protected] and the human gets the link automatically.

Three tiers of memory — because not all memory is equal

A transcript dump isn't memory. Agents need different recall for different things:

# Short-term: ordered, role-aware conversation history (sub-ms reads)agmry store my-project "Deployed v2.1, rolled back — migration locked the users table" -r assistantagmry recall my-project -n 50
# Durable facts: typed key-value context — decisions, preferences, runbooksagmry ctx my-project set deploy_flow "push image to registry, then ask infra to roll"agmry ctx my-project get deploy_flow
# Long-term: titled, tagged knowledge entries that survive monthsagmry entry my-project "Auth decision" "JWT not sessions — mobile clients can't hold cookies" --tags "arch,decisions"agmry entries my-project --tags "decisions"
# Work-in-progress: scratchpad that's allowed to expireagmry scratch my-project set "midway through the billing refactor, invoice.js next"

And semantic search stitches it together when the agent only half-remembers:

agmry search my-project "did we ever discuss rate limiting"# → surfaces a months-old entry, with similarity score

Session start = one command

Instead of pasting yesterday's summary into today's prompt:

agmry boot my-project --json          # messages + context + entries, one callagmry boot my-project --semantic "billing refactor"   # or focused on a topic

~200 tokens of structured state, not top-k chunks of old transcripts.

MCP Server

Prefer tools over a CLI? agmry ships an MCP server. Point Claude Code (or any MCP client) at it and your agent gets 17 native tools: store, recall, search, bootstrap, context, entries, spaces, queues.

claude mcp add agenticmemory -- agmry mcp-serve

[MCP in Claude Code — Claude remembers across sessions]

Claude Code remembering across sessions via MCP — click to watch.

For clients that use a JSON config (Cline, Cursor, Windsurf), pass your API key via the environment — the MCP server runs outside your project directory, so it won't pick up .agmry/config.json:

{  "mcpServers": {    "agenticmemory": {      "command": "agmry",      "args": ["mcp-serve"],      "env": { "AGMRY_API_KEY": "amk_your_key_here" }    }  }}

Remote MCP — zero install

Claude Web, Claude Desktop, Raycast, or any hosted MCP client can connect straight to the remote server. Same tools, same API key, nothing to install:

URL:  https://mcp.agenticmemory.ai/sseAuth: Authorization: Bearer YOUR_API_KEY

End-to-end encryption (zero-knowledge spaces)

Create a space the server can never read. Encryption happens inside the CLI — the API only ever sees ciphertext.

agmry key generate                                  # one-time: creates + saves your keyagmry space create "Private" private --encryption e2e
agmry store SPACE "my secret"                       # encrypted before it leaves your machineagmry recall SPACE                                  # transparently decryptedagmry key verify SPACE                              # holding the right key?

Or derive per-space keys from a passphrase instead of storing a key:

agmry key set --passphrase "long secret phrase"

Notes:

• Semantic search is impossible on zero-knowledge spaces by design.
• Prefer encrypted at rest instead? --encryption managed — the server encrypts your data at rest and every feature (search, summaries) keeps working. No client key needed.
• Losing the key or passphrase means the data is unrecoverable. That's the point. Back it up: agmry key show --reveal.

Key resolution order: --enc-key / --passphrase flag → AGMRY_ENCRYPTION_KEY env → ./.agmry/config.json → ~/.agmry/config.json. Same envelope and key derivation as the Node and Python SDKs — keys are interchangeable across all three.

Multi-agent: one brain, many agents

Spaces are shareable. One agent stores the deploy runbook; another recalls it a week later, from a different machine, over a different interface (CLI, MCP, or REST — same memory). Your agents hand off between sessions and between projects without you couriering context.

And you stay in the loop: everything your agents remember is browsable in a human dashboard.

[Human dashboard — everything your agents remember]

Everything your agents remember, in one dashboard — click to watch.

Queues — the agent bus

FIFO queues inside a space. One agent pushes work, another pops it — no polling glue, no extra infra.

agmry queue SPACE jobs push '{"task":"review PR #42"}'   # enqueue (FIFO)agmry queue SPACE jobs pop                                # dequeue oldest — exit code 2 if emptyagmry queue SPACE jobs pop --wait 25                      # long-poll up to 25s for the next itemagmry queue SPACE jobs                                    # peek: length + head, without consumingagmry queue SPACE jobs dlq push '{"task":"..."}' --reason "failed twice"   # dead-letteragmry queue SPACE jobs dlq list                           # inspect dead-lettered items

Exit code 2 on empty means shell loops branch cleanly: while agmry queue SPACE jobs pop --wait 25 --json; do ...; done. Envelopes are opaque JSON — the server never inspects them.

Agent Integration

Add to your CLAUDE.md, .cursorrules, .clinerules, .windsurfrules, or AGENTS.md:

## Agentic MemoryThis project uses Agentic Memory for persistent memory across sessions.Use the `agmry` CLI. Key is in .agmry/config.json (auto-loaded).
agmry boot SPACE --json               # load everything at session startagmry store SPACE "what happened"     # remember somethingagmry ctx SPACE set key "value"       # store a durable decision/factagmry search SPACE "the deadline"     # find past contextagmry queue SPACE jobs push '{...}'   # send work to another agentagmry queue SPACE jobs pop --wait 25  # receive work (exit 2 = empty)

Config Priority

1. --key flag
2. AGMRY_API_KEY environment variable
3. AGENTICMEMORY_API_KEY environment variable
4. ./.agmry/config.json (project-local)
5. ~/.agmry/config.json (global)

Add .agmry/ to your .gitignore.

Features

• Conversation history — ordered, role-aware messages with recency windowing, sub-ms reads
• Key-value context — typed durable facts: decisions, preferences, runbooks
• Long-term entries — titled, tagged knowledge that survives months, with auto-summarisation
• Entities — people and systems the agent should know about
• Scratchpad — ephemeral working memory with TTLs (expiry is a feature)
• Semantic search — across everything the agent has ever stored
• Bootstrap — full session context in one call
• Agent self-signup — working API key from one CLI command, live for 48h keyless; $0 card-auth starts the free 7-day trial
• Queues — FIFO agent bus with long-poll and dead-letter, agmry queue (exit 2 = empty)
• MCP server — 46 tools, local (agmry mcp-serve) or fully remote (mcp.agenticmemory.ai)
• REST API — same memory on the request path of proxies and pipelines
• Multi-agent spaces — a fleet of agents reads and writes one memory
• End-to-end encryption — zero-knowledge spaces where only you hold the key (agmry key generate)
• Export — full data takeout per space, one command (agmry space export)

Pricing: The agent's key works instantly — full access for 48h, no card. A $0 card-auth starts the free 7-day trial; after that from $24.99/mo. No data deletion, ever: your memory waits for you. Details.

Why not just Claude's built-in memory?

Claude Memory is great — but it only works with Claude, and only Claude decides what's in it.

Why this exists

I was my agents' memory. Every session started with me re-explaining my own project to my own tools — decisions, preferences, what broke last time. So I built the memory they run themselves. My own agents use it in production across a dozen projects; if something's rough or missing, open an issue — I read every one.

Documentation

• Quickstart guides — Claude Code, Cursor, LangChain, CrewAI, AutoGen + 7 more
• API docs
• Pricing

License

Proprietary — Tyga.Cloud Ltd. See LICENSE.