Introduction
Keep durable AI project knowledge in a local, searchable, git-friendly store.memorystore is a small MCP server for project-local AI memory. It gives agents a durable place for verified setup facts, architecture decisions, debugging outcomes, and repository conventions without putting that knowledge in a global account or scattering scratch files through the project.
.memory-store/
metadata.toml
VERSION
.gitignore
memories/
<cuid2>.md
.cache/
fff/Markdown files are the canonical store. TOML frontmatter carries the generated key, tags, timestamps, and schema version, while an FFF-backed cache makes the note bodies searchable. The cache is disposable and ignored by Git.
The MCP interface exposes six operations: initialize the store, create a note, search, read by key, update with exact edits, and delete with revision checking. Agents never need direct filesystem access to .memory-store.
What belongs here
Memory Store works best for knowledge that should survive across AI sessions and remains specific to one project:
- A setup command that was verified on the repository.
- A design decision with a short reason.
- A bug symptom and its confirmed root cause.
- A convention future agents should follow.
- A completed feature note explaining what changed and why.
Committed documentation is still the right place for user-facing guides and broad architecture. Memory Store fills the smaller operational gap between source code and transient chat context.