Client
Editor hooks, fail-open
Recall injects before the prompt. Capture queues after the session. If Brainfeather is slow, the editor continues.
Long-term memory for AI agents
Your agent forgets
everything. We don't.
Brainfeather is the memory layer that sits under Claude Code, Cursor and your own agents — recording the facts that matter and handing them back on the next run.
One store, every client
Write once. Recall everywhere.
Facts captured in one tool are available in all of them. Change a decision in Cursor and Claude Code knows about it on its next run.
MCP
stdio or /mcp
Pin @brainfeather/mcp@1.6.1. Context is compiled on the Singapore API so the model does not wait on a second extraction call.
Dashboard
Review, then recall
Inferred facts wait at /review. Approved memories are what the next Cursor, Claude Code, or OpenCode session sees.
In the prompt
What the model sees
Same model, same question, same day's work — the only difference is whether it already knows the relevant context. The coding example below is one case; Brainfeather can also remember how you communicate, learn, plan, and make decisions.
claude — ~/projects/my-appno mcp
i'm creating a new app — set up the backend
Before I scaffold anything — a few questions:
Which backend — Supabase, Firebase, something custom?
Managed auth, or your own?
Where should generated files live?
…you re-type your whole stack. Again.
Glob(pattern: "**/*.config.*")
No files found
Write(src/data/repository.ts)
Added 47 lines
1import { Orm } from '@orm/client'
2
3export class DataRepositoryService {
4 private orm = new Orm()
… +43 lines (ctrl+o to expand)
✗ an ORM, a service class, a folder you don't use
Done(5.2s · ↓ 1.5k tokens)4 turns
38% context used
claude — ~/projects/my-appmcp: brainfeather
i'm creating a new app — set up the backend
brainfeather - search_memory(query: "backend setup")
Found 4 facts
- backend
- Supabase · Postgres
- auth
- Supabase Auth + RLS
- testing
- Vitest · colocated *.test.ts
- convention
- thin repositories, no ORM
Your Supabase client, thin repository:
Write(src/repositories/user.ts)
Added 24 lines
1import { supabase } from '@/lib/supabase'
2
3export const userRepo = {
4 findById: (id: string) =>
5 supabase.from('users')
6 .select().eq('id', id).single(),
7}
… +17 lines (ctrl+o to expand)
✓ your provider, your conventions, your file layout
Done(2.6s · ↓ 489 tokens)1 turn
12% context used
Beyond project memory
It remembers your work.
Then it learns how you work.
Brainfeather can hold your stack and decisions, but personalization goes further: preferences, goals, routines, constraints, and the way you want an assistant to respond.
brainfeather.config
1
2
3
4
5
6
2
3
4
5
6
{
"mcpServers": {
"brainfeather": {
"command": "npx",
"args": ["-y", "@brainfeather/mcp@1.6.1"]
}
}
}Claude Code
Cursor
OpenCode
01 / Connect
One setup. Every coding agent remembers.
Add Brainfeather as an MCP server once, then run init so Cursor, Claude Code, and OpenCode recall automatically. Client → MCP → dashboard: inferred facts wait in review until you approve them.
I understand things faster when I see one concrete example first.
Keep the first answer short. I'll ask when I want the deep version.
My current goal is to become better at systems thinking and leadership.
Understanding
How you work
communication
Concise first, detail when asked
learning
Examples before theory
goal
Grow into a systems leadership role
constraint
Deep work after 6 PM
02 / Understand
It learns the person behind the prompt.
From what you explicitly share, Brainfeather remembers how you prefer to communicate, how you learn, what you are working toward, and the constraints shaping your day.
Living context
One memory folder
Preference
Concise first
Goal
Systems leadership
Schedule
Deep work after 6 PM
Decision
Current context wins
Session 1
context fills
Compaction
chat shrinks
Session 2
context restored
One living folder · current in every connected agent
03 / Continue
Stop rebuilding context files every time something changes.
Connected agents can write durable updates through MCP instead of repeatedly copying live context into AGENTS.md, CLAUDE.md, or client-specific rules. The relevant memory remains available after compaction and across new sessions.
You
Help me plan this week so I can make progress without burning out.
concise firstexamples helpevenings freesystems goalprotect deep work
Personalized response
Keep the plan light: three 45-minute evening sessions. Start each with one visual example, then build a tiny system-design exercise. Leave Wednesday open so you do not lose your deep-work rhythm.
your pace·your style·your goals
04 / Personalize
The next response fits how you think and work.
Before answering, your assistant recalls the relevant parts of your working style, goals, schedule, relationships, and project context — then responds in a way that feels made for you.
Questions, answered
How persistent memory actually works.
Do I still need AGENTS.md or CLAUDE.md?
Static instruction files can still be useful for fixed repository rules. Brainfeather removes the need to keep copying changing context, preferences, goals, and decisions into every client-specific file. Connected agents save those durable updates through MCP and recall them when relevant.
What happens when the context window fills up or gets compacted?
The chat context may shrink, but Brainfeather memory lives outside the conversation. A connected agent can retrieve the relevant facts again in the same project, after compaction, or in a new session.
Does Brainfeather silently remember everything I do?
No. Brainfeather is explicit and selective. Connected agents call the MCP tools to save durable facts, and the memory pipeline filters obvious conversational noise. Inferred captures wait at /review and never enter recall until you approve them. You can inspect, correct, retract, or delete what was stored.
Is it only for coding and project context?
No. Project memory is the first use case, but Brainfeather can also hold preferences, goals, routines, constraints, recurring people, and communication style when you explicitly choose to save them.
What happens when something changes?
New corrections can supersede old facts. Brainfeather keeps the history for auditability while active reads return the context that still holds, so stale preferences and decisions do not compete with the current truth.
Which assistants can use the same memory?
Any MCP-compatible client can connect to the same Brainfeather account. Today that includes tools such as Claude Code, Cursor, OpenCode, and your own agents.
Stop re-explaining yourself.
Brainfeather is in early development and onboarding testers. Leave your email and we'll invite you to build an assistant that remembers both your work and the way you work.