AgentRecall-X

# AgentRecall

> One memory shared by every AI agent you use. Rules, corrections and project state follow the user from Claude Code to Codex to Cursor and any MCP client. Use the free hosted server (https://agentrecallx.com/mcp, open sign-up with GitHub) or run it locally and fully offline (npx -y agent-recall-mcp, plain markdown in ~/.agent-recall/).

If you are an AI agent and AgentRecall's tools are connected, run this loop without being asked. Saying "noted" is not saving: only a tool call saves.

- Session begins: call `session_start` first, before you act (`project: "auto"` unless told otherwise). It returns the owner card, the HARD RULES and the project state card.
- A fact or decision worth keeping: call `remember` with `content`.
- The human corrects you: call `check` with `human_correction: { rule, why, applies_when }`. `rule` is one imperative sentence, `why` is the evidence, `applies_when` is 1-5 topic keywords. On the hosted server this is staged until the owner approves it; tell the human it is waiting.
- Before publish, push, deploy, delete or any irreversible action: call `check` with `action_description`. If the `verdict` is `blocked`, stop and read the correction.
- Past context would help: call `recall` with `query`. Safe any time.
- The human says "save" or "remember", or you are about to stop: call `session_end` with `summary`, optional `insights`, and optional `state` (goal, open_decisions, blockers, next, pointers).
- Don't claim you will remember something without calling a tool. Don't paste the whole memory back. Don't store secrets, tokens or credentials.

## Connect

- Instructions written for agents: https://agentrecallx.com/connect.md (hosted) and https://agentrecallx.com/local.md (local); the same index is served at https://agentrecallx.com/llms.txt
- Humans: https://agentrecallx.com/ (sign in with GitHub, get an ID, copy a command)

- Claude Code, hosted: `claude mcp add --scope user --transport http agent-recall https://agentrecallx.com/mcp`, then `/mcp` → Authenticate
- Codex, hosted: `codex mcp add agent-recall --url https://agentrecallx.com/mcp` (re-run sign-in with `codex mcp login agent-recall`)
- Cursor, hosted: `~/.cursor/mcp.json` → `{ "mcpServers": { "agent-recall": { "url": "https://agentrecallx.com/mcp" } } }`
- claude.ai: add a custom connector with the URL `https://agentrecallx.com/mcp`
- Claude Code, local: `claude mcp add --scope user agent-recall -- npx -y agent-recall-mcp`
- Codex, local: `codex mcp add agent-recall -- npx -y agent-recall-mcp`
- Other clients, local: `{ "mcpServers": { "agent-recall": { "command": "npx", "args": ["-y", "agent-recall-mcp"] } } }`

## Tools

- `session_start`: load context; `mode` is `cards` (default), `full` or `lite`
- `remember`: save one fact or decision
- `recall`: ranked search across all memory, with `feedback` to re-weight results
- `check`: record a correction, or check a risky action before doing it
- `session_end`: save the session journal, insights and the project state card
- With `--full` (local, or a hosted server whose operator enables it): `check_action`. Local only, with `--full` and `AR_EXTRAS=1`: `pipeline_open`, `pipeline_close`, `pipeline_list`, `pipeline_current`, `pipeline_show`, `register_rule`, `digest`

## Docs

- [README](https://github.com/Goldentrii/AgentRecall-X/blob/main/README.md): overview, install, architecture, the math
- [Hosted server](https://github.com/Goldentrii/AgentRecall-X/blob/main/docs/hosted-server.md): limits (100 MB, 120 requests/min per user), the correction approval gate, running your own server
- [Privacy policy](https://agentrecallx.com/privacy): hosted memory is private per user and stored in the EU (source: [PRIVACY.md](https://github.com/Goldentrii/AgentRecall-X/blob/main/PRIVACY.md))
- [Support](https://agentrecallx.com/support): contact, GitHub issues, best-effort response time
- [AGENTS.md](https://github.com/Goldentrii/AgentRecall-X/blob/main/AGENTS.md): full lifecycle doctrine and CLI equivalents
- [Full reference](https://github.com/Goldentrii/AgentRecall-X/blob/main/README.full.md): SDK API, CLI, every tool parameter