Agentmem

What is Agentmem?

Agentmem is a shared memory system for Claude Code, Cursor, and Codex that tracks the trustworthiness of stored information. It saves sessions, detects stale and conflicting rules, and prevents AI coding assistants from repeating old mistakes. Agentmem is built for developers who want their agent to remember what is still true across sessions.

How to use Agentmem?

Install Agentmem with pip install quilmem[mcp], then run agentmem init --tool claude --project myapp to set up. Restart your editor to give your agent 13 memory tools. For Claude Code, add the MCP server configuration to .claude/settings.json. Agentmem also provides a Python API and CLI for direct use.

Key features of Agentmem

• Memory lifecycle states: hypothesis → active → validated → deprecated → superseded
• Built-in conflict detection with sentence-level negation matching
• Staleness detection by age, source file missing, or hash drift
• Health scoring system (0–100) for memory system trustworthiness
• Trust‑ranked recall: validated memories rank above active and hypothesis
• Local‑first, zero‑infrastructure: SQLite with FTS5 search, no cloud needed

Use cases of Agentmem

• Stop your AI assistant from repeating the same bug fix across sessions
• Catch outdated rules when source files change or age out
• Save and restore session context so the agent picks up where it left off
• Promote confirmed decisions and deprecate disproven hypotheses
• Sync canonical markdown files into the database with provenance tracking

FAQ from Agentmem

What makes Agentmem different from other memory tools?

Agentmem focuses on trust, not just storage. It uses memory lifecycle states, conflict detection, staleness detection, and health scoring to ensure the agent retrieves only what is still true.

What are the runtime dependencies?

Agentmem requires only SQLite with WAL mode. No external databases, API keys, or cloud infrastructure are needed. It uses FTS5 for search and porter stemming.

How does recall work?

Recall uses a composite score of text relevance (25%), trust status (20%), provenance (20%), recency (15%), frequency (10%), and confidence (10%). Deprecated and superseded memories are excluded.

Can I use Agentmem with multiple projects?

Yes. You can scope memories by project using the project parameter in the Python API or CLI. Each project keeps its own set of memories in the same database file.

What license is Agentmem released under?

Agentmem is released under the MIT license.