Store and retrieve agent knowledge with persistent memory

agentMemory gives an agent a persistent, searchable memory bank that automatically syncs with a project's own Markdown documentation, run via an MCP server. Setup checks for an existing install (ls -la .agentMemory), then runs npm install, npm run compile, and npm run start-server <project_id> <absolute_path_to_workspace> to bring the server up as a background or MCP-configured process; Node.js has to be installed first. Once running, it exposes four tools: memory_search (search by query, type, or tags - for example memory_search({ query: "authentication", type: "pattern" })), memory_write (record new knowledge or a decision under a key, type, and content, with optional tags), memory_read (retrieve one memory's full content by key), and memory_stats (usage analytics on the memory bank). On first run in a project it can import existing Markdown memory banks from .kilocode/, .clinerules/, or .roo/ rather than starting from an empty store.

Use it whenever a hybrid, persistent, searchable memory system is needed for an AI agent working in a project. The intended loop is to search memory for relevant context before a task, use read or search to answer questions while the task is in progress, and write new findings back to memory once it is done, with every write automatically synced to standard Markdown files rather than staying locked in a private database only the agent can read.

Input to memory_write is a key, a type, content, and optional tags; input to memory_search is a query plus optional type or tag filters; input to memory_read is just the key. Output is the matched memory record or records for search and read, a confirmation for a write, and a statistics summary from memory_stats covering how the memory bank is actually being used. Every write is also mirrored into the project's own Markdown files, so the memory bank stays readable outside the tool by anyone browsing the repo directly.

An MCP server (npm run start-server) is required to expose the four memory tools. It reads from and syncs to plain Markdown files, including pre-existing memory banks under .kilocode/, .clinerules/, or .roo/ if those already exist in the project.

Teams running AI agents across long-lived projects who want architecture decisions, patterns, and other findings to persist and stay searchable across sessions, instead of being re-derived or re-explained to the agent every time, with the same knowledge also readable as plain Markdown by any human on the team.