Agent Memory — 错题本

The error notebook for AI agent tool calls.
pip install agent-memory → one SQLite file → your agent stops repeating mistakes.

Why

Every AI agent hits the same wall: it calls web_search, gets a timeout, retries with the same long query, times out again. With Agent Memory, it learns: "long queries cause timeouts → shorten them." Next time, it gets it right on the first try.

Benchmark (3 tasks × 5 learning rounds):

Memory doesn't just store data — it stores lessons. The agent re-discovers fewer fixes, wastes fewer tokens, and completes tasks faster.

Install

pip install agent-memory

# With optional semantic search:
pip install agent-memory[embed]

# With MCP server:
pip install agent-memory[mcp]

Quick Start

from agent_memory import AgentMemory

mem = AgentMemory("./agent_errors.db")

# Log a tool call outcome
mem.remember(
    tool="web_search",
    input="GRPO reinforcement learning survey with comparison...",
    outcome="fail",
    error="timeout",
    error_detail="Request timed out after 30s",
    fix="shorten query to under 80 chars",
)

# Search relevant past experiences
results = mem.recall("web_search timeout")
for r in results:
    print(f"[{r['outcome']}] {r['tool']}: {r['fix']} (confidence: {r['confidence']:.0%})")

# Review memories for manual correction
mem.review(low_confidence_only=True)

# After using a memory, confirm if it helped
mem.confirm(memory_id=1, was_helpful=True)

MCP Server

Use as an MCP server that any agent framework can connect to:

agent-memory serve --db ./agent_errors.db

Add to your MCP client config:

{
  "mcpServers": {
    "agent-memory": {
      "command": "agent-memory",
      "args": ["serve", "--db", "./agent_errors.db"]
    }
  }
}

Available MCP tools: remember, recall, review, forget, confirm, stats.

How It Works

1. Remember: Agent logs every tool call — success or failure, with error type and fix
2. Store: SQLite with FTS5 trigram tokenizer (English + CJK), time-decay ranking, confidence scoring
3. Recall: Agent queries past experiences before making decisions
4. Learn: Confidence adjusts based on whether the memory actually helped

No LLM calls — all storage and retrieval is pure rules + SQL. Your agent's LLM only sees the relevant memories as context.

Write Strategy

Anti-Pollution

• Confidence scoring: +0.2 when memory helps, -0.2 when it misleads
• Time decay: Older memories rank lower (configurable half-life, default 30 days)
• Human review: review() for manual inspection and forget() for removal

Roadmap

License

MIT