The failure mode of agent memory is that recalling it costs more context than it saves. Returning a table of section names, sizes and staleness — and no content — turns the read into a decision rather than a dump, and the agent then asks for the two sections it actually needs. The staleness and size warnings work the same way: they surface the sections that have quietly rotted, which is the thing nobody notices until the memory is actively wrong.
A persistent memory layer for AI coding agents, stored as local markdown files and shared across any IDE that speaks MCP. Project context is written once and read back everywhere. Projects are recognised by path or by git remote URL, so moving a folder or switching machines does not lose the memory.
- A project registered or reconnected, returning a memory summary with warnings — empty sections, sections stale beyond 7 days, sections over 10k chars, and content older than 30 days that suggests pruning — `init_project`
- Memory read in the mode you need: a compact summary table by default with no content at all, only named sections, a character cap per section, a keyword search across sections, previous versions, or everything with cleanup instructions — `read_memory`
- Sections written or appended, with the previous content saved to history automatically; append mode adds timestamped entries without rewriting — `write_memory`
- Projects listed or deleted — `manage_projects`
- Three prompt templates that tell the agent when to use memory at all: `start_session` at the beginning of a conversation, `bootstrap_memory` for learning an existing project, `update_memory` at the end of a productive session
- Default sections for every project — `overview`, `decisions`, `active_context`, `progress` — plus any custom ones you add, with the last 5 snapshots kept per section
No account, no database, no cloud — local markdown files. Python 3.11+; install with `uv pip install ide-memory-mcp` or pip, then run `ide-memory-mcp setup` to detect and configure installed IDEs (Cursor, VS Code, Windsurf, Claude Desktop) and restart the IDE. Manual configuration is a one-line `command` entry in the IDE's MCP config. Two helper commands come with it: `doctor` verifies the install, storage use and which IDEs are configured, and `status` lists projects with section counts, size and last update. The pyproject name is `ide-memory-mcp` (1.0.0), MIT licensed.
One command — pip install ide-memory-mcp