MCP 服务器

Recall

Open-source MCP memory server for AI agents — persistent, searchable, tiered memory across sessions. Works over stdio (Cursor, Claude Desktop) or HTTP+SSE. MIT licensed.

README

<div align="center">

Recall™

A better memory server for AI agents — works for one, scales to many. Local, free, zero-config, MCP-native. Your data stays on your machine.

Quickstart · vs. mem0/Letta/Zep · Multi-agent · Recall Pro → · Book a demo

</div>

Without a memory server, every Claude / Copilot / Cursor conversation starts cold. You re-explain the codebase, the conventions, the decisions, the gotchas — every time. Recall fixes that.

Start here: what Recall does for one developer, one AI

Install it once, point your MCP client at it, and your AI now:

Remembers across sessions — "what did we decide about the auth flow last week?" returns the actual decision, not a hallucination
Indexes your code and docs — index_file + recall = local semantic RAG over your repo
Cites where the answer came from — answer returns text plus the chunks it pulled from
Builds project knowledge — every checkpoint, reflect, and anti_pattern becomes searchable later
Survives restarts — append-only artifacts on disk, vector store rebuildable from them

One pip install, one config block, done. No API key. No external service. No per-token bill. MIT license. This is what 95% of users will ever use Recall for.

How is this different from mem0 / Letta / Zep?

Recall does the same job they do — persistent memory across AI sessions, semantic recall, "remember what the user said last week." The difference is where and how:

	mem0 / Letta / Zep	Recall
Where memory lives	Their cloud	Your `~/.recall/`
API key required	Yes	No
Cost	Per-token / monthly SaaS	Free
Embeddings	Their service	Local ONNX (offline)
Network calls	Every recall	Zero
Air-gappable	No	Yes
MCP-native	Wrapper or SDK	Built on MCP
Multi-agent coordination	None	6 primitives

If you're happy paying a hosted memory provider per token, those are great products and you don't need Recall. If you'd rather your AI's memory live on your laptop or your own server, free and offline, that's what Recall is for.

Scaling up: coordination when you run more than one agent

The same install that gives one developer a personal AI memory also works as a shared brain when more than one agent talks to it. Two Copilot windows. A planner + executor pair. Three Claude instances dividing up a refactor. A pre-commit agent and a code-review agent on the same PR. They all remember and recall from the same store.

That introduces a new problem none of the hosted memory services have even tried to solve: agents stepping on each other. Agent A starts refactoring src/auth.py. Agent B, in another window, rewrites the same file with no idea A is mid-edit. Whoever saves last wins. The other agent's work is gone.

Recall ships six MCP primitives that turn parallel agents from a clobber-fest into a coordinated team:

Tool	What it does
`claim(resource, agent)`	Soft-lock a file/table/URL with an auto-expiring TTL
`release(resource, agent)`	Drop the lock (soft-archive — audit trail survives)
`who_has(resource)`	"Is anyone editing `src/foo.py` right now?"
`claims()`	All active locks across all agents
`handoff(to_agent, ...)`	Explicit work transfer with intent + files + context
`pulse_others(self_agent)`	The N most recent checkpoints from agents other than you

Claims are advisory (like git locks) — Recall doesn't physically stop a second agent from writing, but every well-behaved client checks first. TTLs prevent a crashed agent from freezing a resource forever. Releases soft-archive (per the project-wide delete=archive rule) so the audit trail of who held what when survives.

If you're a single user, these tools just sit there unused. If you ever scale up to multiple agents, they're already there.

   ┌──────────────┐                           ┌──────────────┐
   │  Agent a3f7  │      claim(file, ttl)     │  Agent b1c4  │
   │  Claude #1   │ ───────────┐  ┌─────────► │  Claude #2   │
   └──────┬───────┘            ▼  │           └──────┬───────┘
          │              ┌────────┴───────┐          │
          │   remember   │     Recall     │   pulse  │
          ├────────────► │ • shared memory│ ◄────────┤
          │              │ • claims/locks │          │
          │   handoff    │ • handoffs     │  handoff │
          ├────────────► │ • who_has      │ ◄────────┤
          │              └────────────────┘          │
          ▼                                          ▼
       22 MCP tools — Copilot, Claude, Cursor, custom

22 MCP tools total — 16 memory tools every user gets, plus the 6 coordination primitives that activate when you scale up.

One-line install (Claude Desktop, VS Code, Cursor)

Recall ships as a stdio MCP server. Zero config — no API keys, no Docker, no ports. Memory lives in ~/.recall/.

pip install "ai-recallworks[mcp]"

Then add Recall to your MCP client config:

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "recall": {
      "command": "recall-mcp"
    }
  }
}

VS Code (mcp.json in your workspace or user settings):

{
  "servers": {
    "recall": {
      "command": "recall-mcp"
    }
  }
}

Restart the client. Your agent now has persistent memory across sessions. Embeddings run fully offline (Chroma's bundled all-MiniLM-L6-v2). Upgrade to Ollama / OpenAI / Voyage embeddings via env vars when you want.

Five-minute install (HTTP / multi-user / team)

1. Run the server:

docker run -d --name recall \
  -p 8787:8787 \
  -e API_KEY=changeme \
  -v recall-data:/data \
  ghcr.io/recallworks/recall:latest

2. Talk to it — pick your stack:

# Raw HTTP (any language)
curl -H "X-API-Key: changeme" \
     -H "Content-Type: application/json" \
     -d '{"content":"first memory","tags":"hello"}' \
     http://localhost:8787/tool/remember

# Python (use requests/httpx — no SDK pkg needed)
import requests
h = {"X-API-Key": "changeme", "Content-Type": "application/json"}
requests.post("http://localhost:8787/tool/remember", headers=h,
              json={"content": "first memory", "tags": "hello"})
print(requests.post("http://localhost:8787/tool/recall", headers=h,
                    json={"query": "memory"}).json()["result"])

// TypeScript / JavaScript (Node 18+, Bun, Deno, browser)
npm install @recallworks/recall-client

import { RecallClient } from "@recallworks/recall-client";
const c = new RecallClient({ baseUrl: "http://localhost:8787", apiKey: "changeme" });
await c.remember("first memory", { tags: "hello" });
console.log((await c.recall("memory")).result);

Full walkthrough: docs/quickstart.md.

What you get

13 tools — remember, recall, reflect, anti_pattern, checkpoint, pulse, session_close, index_file, reindex, snapshot_index, memory_stats, forget, maintenance.
Two transports — plain HTTP (POST /tool/{name}) and MCP over SSE. Drop into Copilot, Claude Code, Cursor, or any MCP client.
Bring your own models — pluggable embedder (default / OpenAI / Ollama) and summarizer (noop / OpenAI / Ollama). Run fully offline, fully on-prem, or against your own Azure-OpenAI tenant. See docs/byo-models.md.
Durable by default — ephemeral live store with auto-snapshot to disk; container restarts come up whole.
Append-only artifacts — every write also lands as a .md file. If the vector store ever burns down, reindex rebuilds it from the artifacts.
forget is soft-archive — guardrail wired into the OSS code itself, not bolted on as policy. Memory you delete can be recovered.

How it's different

	Recall	Mem0 / Letta / Zep
License (core)	MIT	mixed; SaaS-first
Self-host	one `docker run`	varies, often non-trivial
BYO embedder	default / OpenAI / Ollama (env var)	usually fixed
BYO LLM	noop / OpenAI / Ollama (env var)	usually fixed
Storage model	append-only artifacts + vector index, rebuildable	live DB only
`delete`	soft-archive by design	hard delete
Tool surface	13 opinionated tools (memory + workflow)	embedding + retrieval primitives
MCP-native	yes, plus plain HTTP	partial / via wrapper
Ops model	single binary, single container	multi-service stack

If you want a managed service, see Recall Cloud below. If you want a brain you fully own, this OSS core is enough.

Repo layout

Path	What
`src/recall/`	OSS server (MIT)
`src/recall/tools/`	One module per tool
`src/recall/transport/`	HTTP + MCP/SSE adapters
`docker/single-tenant/`	Reference Dockerfile + compose
`tests/`	pytest suite (no Docker required)
`docs/`	Quickstart, conventions, architecture
`enterprise/`	Multi-tenant, SSO, control plane (BSL)

Conventions

These are the practices that make the tools pay off. Pick what fits.

Cold-start ritual — opening protocol every session should run.
Branding — signed-edit headers so you can trace which agent touched which file when.

Status

Alpha. The code in src/recall/ is extracted from a hosted production brain that has served thousands of sessions, then sanitized of org-specific paths, extensions, and tenant data. Expect breaking changes before 1.0; pin the image tag.

Contributing

Yes — please read CONTRIBUTING.md first. We accept bug fixes, new Store backends, doc improvements, and anti-pattern entries. We don't accept architectural rewrites without prior discussion.

Security issues: see SECURITY.md.

License

src/recall/, clients/, docker/single-tenant/, docs/, examples/ — MIT (LICENSE)
enterprise/ — BSL 1.1, 5-seat additional-use grant, converts to MIT after 3 years (LICENSE-COMMERCIAL.md)

Recall Open Source vs. Recall Pro vs. Hosted

Capability	OSS (this repo)	Recall Pro	Recall Cloud
Single-tenant Docker image	✅	✅	n/a (hosted)
13 memory tools, MCP + HTTP	✅	✅	✅
BYO embedder + LLM	✅	✅	✅
Append-only artifacts + auto-snapshot	✅	✅	✅
Multi-tenant, SSO, RBAC	—	✅	✅
Audit log + retention policy	—	✅	✅
Cross-session entity graph	—	✅	✅
PII sanitization pipeline	—	✅	✅
Snapshot replication / DR	—	✅	✅
Vendor support + SLA	community	business hours	24×7
Hosted on our infra	—	—	✅
Pricing	free	from $99/mo per node	from $0.10 per 1k tools

Recall Pro ships from the enterprise/ tree under a Business Source License — source-available, 5-seat free Additional Use Grant, converts to MIT after 3 years. Buy a license and the enterprise/ modules light up alongside your OSS install.

Recall Cloud is the hosted multi-tenant version. Same tools, no infra. Reach out for early-access pricing.

➡️ Talk to sales: sales@recall.works · Book a 20-min walkthrough: https://recall.works/demo

Vertical builds powered by Recall

Recall is the engine. We ship turn-key vertical brains on top of it:

IceWhisperer — the memory + workflow brain for ICE Mortgage Technology / Encompass shops. Pre-loaded SDK index, settings recipes, plugin audits, drift detection. Pilots from $250/mo.

If you want a vertical brain for your industry, we'll build it. Email partners@recall.works.

Maintainers

Reach the maintainers at maintainers@recall.works. Issues and PRs welcome on GitHub.