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:

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:

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