Guide
AI agent memory for coding agents
A coding agent starts every session knowing nothing about your project. It relearns the stack, rediscovers the conventions, and asks again about decisions you already made. Agent memory is the layer that fixes this: a durable store the agent reads before it works and writes to as it learns.
A context window is not memory
The context window is working memory. It is fast, it holds one session, and it is compacted or cleared when that session ends. Long-term memory lives outside it. It survives restarts, it is shared between tools, and it has to answer a question a transcript never does: which of these facts is still true?
What a memory layer has to get right
Keep what lasts
Decisions, conventions and constraints, not greetings or thinking out loud.
Retire what changed
A new decision replaces the old one instead of competing with it.
Stay in scope
Facts from one repository, branch or task do not leak into another.
Decline when unsure
Nothing relevant stored should mean nothing returned, not the nearest guess.
Memory over MCP
The Model Context Protocol (MCP) gives agents a standard way to call external tools. A memory server exposed over MCP works with every compatible client at once: Claude Code, Cursor, OpenCode and agents you build yourself read and write the same store, so a decision recorded in one is known to the others on their next run.
How Brainfeather does it
Brainfeather is a memory layer for coding agents, served over MCP. Memories are scoped to a repository and can be narrowed to a branch or a task. Inferred facts wait at a review queue and do not enter recall until you approve them. Superseded facts are kept for history but no longer returned.
| Tool | What it does |
|---|---|
| get_context | Called first: the stack, decisions and conventions on record, within a token budget. |
| search_memory | Looks up a past decision before the agent picks a library or pattern. |
| save_memory | Records one durable fact you stated or confirmed. Never a guess. |
| capture_activity | Queues inferred facts for review. They stay out of recall until approved. |
| forget_memory | Deletes a memory you say was recorded in error. |
| list_entities | Lists the tools, languages and concepts linked to this project's memories. |
| traverse_graph | Shows the memories and entities connected to one entity. |
Testers connect with an API key and run npx -y @brainfeather/mcp@1.6.2 init once, which sets up automatic recall in Claude Code, Cursor and OpenCode.
Brainfeather is listed in the official MCP Registry and on Smithery.
Questions
- What is AI agent memory?
- It is the store an agent reads from and writes to outside its context window, so facts learned in one session are available in the next. For a coding agent that means the project's stack, conventions, decisions and the reasons behind them.
- Isn't a large context window enough?
- No. A context window holds one session and is cleared or compacted when it ends. Memory persists across sessions and across tools, and it has to decide what is still true, which a transcript never does.
- How does an agent use memory over MCP?
- The Model Context Protocol lets any compatible client call a memory server's tools. The agent asks for context at the start of a task and saves durable facts as it learns them. One server can serve every MCP client you use.
- What should a good memory layer refuse to do?
- Return a confident wrong answer. When nothing stored is about the question, returning nothing is better than returning the closest unrelated fact, because the agent will act on whatever it is given.
Brainfeather is in early development and free while it is. Request access or get in touch.