M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
A

Agent Recall

Community
Correction-first agent memory: the only open-source system that measures whether your agent actually stops repeating mistakes.
How to install ↓ Compare GitHub source repository ↗
Category
Other #96 of 230
Stars
★ 371 Very popular
Transport
stdio (local process)
Runtime
Node.js
Credentials
Optional API key
License
MIT
Last commit
Tools
5
57FMRS · C

AgentRecall offers a unique correction-first memory approach and an honest measurement framework, but its current self-reported numbers are low (e.g., capture recall 35.3%, transfer recall 0/4), indicating that its utility is still being validated. Suitable for Claude Code users who value auditability and local control, provided they actively follow the session loop.

Strongest · Setup experience 14/20 Weakest · Reliability 8/20

Reliability
8/20
Security and permissions
10/20
Maintenance
12/20
Documentation
13/20
Setup experience
14/20
Why each score
Reliability 8/20
Evidence is limited to the manifest and README; no source, tests, or CI were supplied. The described happy path (session_start/remember/session_end, 5 default tools) is plausible and the design is detailed. Deductions: no executable verification, no committed tests or CI in scope, and the manifest version 3.4.32 conflicts with a README reference to v3.4.37 features, so confidence is reduced.
Security and permissions 10/20
No red lines triggered: no real tokens in install examples, no covert exfiltration indicators, and default local-only storage is disclosed. Optional network paths (Supabase, OpenAI, Telegram) are mentioned, with claims of scrubbed egress and owner-gated rule edits. Deductions: source was not reviewed, so least privilege, confirmation flows for dangerous operations, and actual data-flow boundaries cannot be verified.
Maintenance 12/20
Positive signals: MIT license, versioned npm packages, changelog, docs, 28 open issues, repo not archived, contributing guide. Deductions: no commit history, release dates, dependency-update evidence, or issue-response metrics were provided, and no security-response channel is documented, so full marks are not justified.
Documentation 13/20
README is layered and detailed: install instructions for multiple clients, SDK/CLI examples, architecture, 5-layer model, command reference, benchmark caveats, and links to fuller documentation. Deductions: exact MCP tool parameter schemas and authentication/troubleshooting details are not fully present in the supplied material, and the v3.4.32/v3.4.37 inconsistency is a documentation defect.
Setup experience 14/20
Setup is well documented: copy-paste commands for Claude Code, Cursor, VS Code, Windsurf, Codex, generic JSON, SDK, and CLI, plus a skill install. The steps are few and target mainstream clients. Deductions: no committed CI or tests were provided to verify the path, and no install troubleshooting section is present, so the score stays below the 15 cap.

Static review · not runListed 2026-08-07

Read the FMRS scoring method →

Fit and risk

What it can accessReads local filesWrites / deletes local files

Best for

  • Claude Code users who want their agent to stop repeating mistakes
  • Developers who want honest measurement of memory effectiveness
  • Privacy-conscious users requiring local storage
  • Teams adopting structured memory layers (episodic, semantic, procedural, etc.)

Not for

  • Scenarios requiring real-time or high-throughput memory retrieval
  • Users wanting fully automatic memory management without manual session loop
  • Projects needing cloud sync or multi-user collaboration
  • Users requiring advanced retrieval performance like BM25 indexing

Required permissions

  • Read and write local filesystem (default path ~/.agent-recall/)
  • Execute local commands (e.g., CLI tools)
  • Optional: OpenAI API usage if OPENAI_API_KEY is set for vector search
  • Optional: Supabase mirror for remote sync

Risks and side effects

  • Memory content may contain sensitive information; ensure access control despite local storage
  • The measurement instrument may produce misleading data (e.g., defaulted to "heeded" bias); interpret cautiously
  • The automatic session loop requires user compliance with session start/end protocols; otherwise memory may be incomplete
  • Experimental components (e.g., dream consolidation, recurrence harness) may be unstable
  • Dependency on npm package execution poses supply-chain risk (verify source)

Setup

Before you start

Runtime：Node.js

OPENAI_API_KEY optionalsecret OpenAI API key used only to enable optional vector semantic recall; obtain it from platform.openai.com. Without it, local keyword + RRF retrieval is used.
  1. Install the MCP server: run claude mcp add --scope user agent-recall -- npx -y agent-recall-mcp in Claude Code. For other clients, use the JSON configuration.
  2. In the first message of every new session, run the loop: call session_start to load context, call remember with type "correction" when corrected, and session_end at session end.
  3. Optionally install the SDK (npm install agent-recall-sdk) or CLI (npx agent-recall-cli).
claude_desktop_config.json 
{
  "mcpServers": {
    "agent-recall": {
      "command": "npx",
      "args": [
        "-y",
        "agent-recall-mcp"
      ]
    }
  }
}

Shown for Claude Desktop. Other clients may use a different file or key (VS Code uses "servers") — the configurator below converts it.

.vscode/mcp.json
{
  "servers": {
    "agent-recall": {
      "command": "npx",
      "args": [
        "-y",
        "agent-recall-mcp"
      ]
    }
  }
}

Goes in your project's .vscode/mcp.json (VS Code uses a "servers" key).

Terminal
claude mcp add agent-recall -- npx -y agent-recall-mcp

Run it in a terminal; replace any <…> placeholders with your own values first.

Generate and validate client config

Check that it works

After installing with claude mcp add --scope user agent-recall -- npx -y agent-recall-mcp, the tools session_start, session_end, remember, recall, and memory_query should appear in the client's tool list; calling session_start and getting loaded context confirms the connection.

Troubleshooting

  1. If the MCP server fails to start, check Node.js version and npm connectivity.
  2. If memories are not persisted, ensure you follow the /arstart and /arsave session loop.
  3. If retrieval results are poor, verify that local memory files exist under ~/.agent-recall/projects/<slug>/.
  4. If you encounter login expiration, run `claude login` to restore automatic features like dreaming.
  5. Consult the official docs/ and UPDATE-LOG.md for further debugging.

Things to try

Once connected, you can ask your AI assistant things like:

  • Call session_start to load context for this project and show pending work
  • Remember this correction: always ask me before making assumptions about the stack
  • Use recall to search my past notes about rate limiting
  • Call session_end to compound what was learned this session

Tools 5

session_start read-only
Call at the start of a session to load context.
session_end writes
Call at the end of a session to compound what you learned.
remember writes
Call when the human corrects you, with type "correction".
recall read-only
Search past knowledge.
memory_query read-only
Query memory on demand.

Use cases

AI-assisted coding with persistent memory across sessions
Tracking whether corrections actually change agent behavior
Local-first memory with no cloud dependency
Managing multiple projects with per-project slugs

Supported clients

Claude Desktop
CursorPartial support
WindsurfPartial support
VS CodePartial support
CodexPartial support

Listed from the project's documentation, not tested by this site.

Overview

AgentRecall is a correction-first agent memory system that provides an MCP server, SDK, and CLI. It records every correction you make to your agent as a structured entry with severity, evidence, and outcome tracking, persisting across sessions, projects, and restarts. Its distinctive corrections ledger and measurement harness track whether a correction actually changes agent behavior in later sessions—not just retrieval. All memory is stored locally in Markdown files by default, with an optional Supabase mirror for vector semantic search.

Similar servers

BioMCP 75 · B

One binary. One grammar. Evidence from the biomedical sources you already trust.

★ 659 · Tools 11 Compare with this →
Compare with other MCP servers

Source revision cfdd084f82a0 Data synced 2026-10-19 Read the FMRS scoring method