Neura Market
Goldentrii/AgentRecall logo

Goldentrii/AgentRecall

Free

跨会话的持久化累积记忆系统，使用智能距离协议召回最相关历史记忆。提供 5 个工具：`session_start`、`remember`、`recall`、`check`、`session_end`。`npx agent-recall-mcp`

FreeFree tier
Type
Open Source
Company
Goldentrii
Visit Website GitHub (40 stars)

About Goldentrii/AgentRecall

AgentRecall is an open-source, correction-first persistent memory system designed for AI agents, particularly Claude Code. It provides a governed corrections ledger that records every agent correction with severity, evidence, and outcome tracking, persisting across sessions, projects, and agent restarts. Unlike other memory tools, AgentRecall includes a measurement instrument that tracks whether a correction actually changes agent behavior in later sessions by recording retrieved_count and outcome (heeded or recurred). The system ships as an MCP server, SDK, and CLI, offering five session lifecycle tools: session_start, remember, recall, check, and session_end. Installation is via a single npx command. The project publishes honest benchmarks showing both strengths and limitations of current recall and behavior change.

Key Features

Governed corrections ledger that stores structured records with severity, evidence, and outcome tracking
Measurement instrument that tracks whether corrections change agent behavior across sessions (heed/recur outcome)
Session lifecycle tools: session_start, remember, recall, check, session_end
Persistence across sessions, projects, and agent restarts
Published honest benchmarks on correction recall and heed rates
Available as MCP server (npx agent-recall-mcp), SDK, and CLI
Open source on GitHub with 311 stars and 53 forks

Pros & Cons

Pros
  • Free and open-source memory solution for AI agents
  • Unique focus on measuring actual behavioral change, not just retrieval
  • Simple session lifecycle API (start, remember, recall, check, end)
  • Self-hosting with transparent benchmarks and no vendor lock-in
Cons
  • Correction capture recall is moderate (~35% in dual-blind audit as per project's own benchmarks)
  • Heed rate (behavioral change) measured as 0/3 events in evidence-grounded post-reset test
  • Requires manual invocation of session lifecycle tools; no automatic detection of corrections
  • Primarily designed for Claude Code; compatibility with other agents may require additional integration

Best For

Providing persistent memory for Claude Code agents that learns from user correctionsMeasuring whether AI agents actually stop repeating mistakes after being correctedMaintaining a cross-session correction history for debugging and improving agent behaviorBenchmarking agent learning and memory systems with reproducible metrics

FAQ

What is AgentRecall?
AgentRecall is an open-source MCP server, SDK, and CLI that gives AI agents a governed corrections ledger and a measurement instrument to track whether corrections actually change agent behavior across sessions.
How does AgentRecall measure if a correction was effective?
Each correction accumulates a retrieved_count, and when the agent encounters the same situation again, it records the outcome as 'heeded' (behavior changed) or 'recurred' (same mistake repeated).
What are the five session tools provided by AgentRecall?
The five tools are: session_start (load context at start of session), remember (store a correction with type, severity, etc.), recall (retrieve relevant corrections), check (query memory state), and session_end (finalize and compound learned information).
Is AgentRecall free to use?
Yes, AgentRecall is open-source and free. It can be installed via npx with no paid tiers required.