Agentmem

What Thezenmonster/agentmem does

Thezenmonster/agentmem MCP server provides governed memory for coding-agent workflows. It stores typed records such as decisions and bugs, preserves session notes, and supports search and context-limited recall. Records carry a trust lifecycle: new observations can begin as hypotheses, move to active, become validated, or be deprecated when they no longer apply. Superseded records remain available for history and link to their replacements.

The storage layer uses SQLite with FTS5 full-text search, so the project can run locally without a separate database or hosted service. Recall ranks validated and active records above hypotheses, while deprecated and superseded records are left out of normal recall. This makes it suitable for projects where old implementation rules or previous mistakes could mislead an agent.

How it works

Thezenmonster/agentmem MCP server can be used through its built-in MCP service, Python API, or command-line interface. The MCP tool set includes operations for adding, finding, listing, updating, and deleting memories, along with session persistence and governance actions. A client can promote a record, deprecate it with a reason, or supersede it with a newer record.

Conflict detection compares related memories using topic overlap and sentence-level negation checks. It distinguishes duplicates from contradictions and reports severity based on the records' states. Staleness checks identify records that have aged beyond a chosen threshold, reference missing files, or no longer match a tracked source hash.

Health reporting combines conflict and stale-memory findings with orphaned references, deprecated-memory weight, and the presence of validated records. The result is a score from 0 to 100 with associated details.

Setup and configuration

Install the MCP extra with pip install quilmem[mcp]. The CLI initialization command accepts a tool and project, for example agentmem init --tool claude --project myapp. The MCP service itself runs with agentmem serve.

A stdio configuration can invoke agentmem with serve, plus --db for the SQLite database path and --project for the project name. The README shows this setup in .claude/settings.json and identifies Cursor and Claude Code as supported workflows. Agent-specific instructions can be copied into files such as CLAUDE.md, .cursorrules, or AGENTS.md so the agent knows when to use memory.

The Python API also accepts a database path and project name when creating a Memory instance. Provenance fields can record a source path, section, and content hash.

Tools and capabilities

The documented MCP tools are:

• add_memory, search_memory, and recall_memory
• update_memory, delete_memory, and list_memories
• save_session and load_session
• promote_memory, deprecate_memory, and supersede_memory
• memory_health and memory_conflicts

The sync behavior is hash-aware: an unchanged source is skipped, a changed source updates the related memory, a removed section causes deprecation, and a restored section can reactivate a deprecated memory.

Limitations and notes

The README documents local SQLite and FTS5 storage rather than a hosted or distributed deployment. The package name used for installation is quilmem, while the command-line executable and MCP configuration use agentmem. The available material does not specify support for authentication, remote hosting, or clients beyond the named Claude Code, Cursor, and generic MCP-client workflows.

Getting started with this Thezenmonster/agentmem MCP server

Always refer to the official documentation for the most accurate and up-to-date information.