Open Agentic Brain and Memory

Built by Rohith, Chandan and Sushmita

Claude Code Codex Cursor Antigravity Kilo Codebeta

One memory for all your AI coding tools.

Claude Code, Cursor, Codex and the rest forget everything when a chat ends, and none of them know what the others learned. I remember for all of them.

Built by Rohith, Chandan and Sushmita

Scroll

Tell one tool. Every tool knows.

Decide something in Claude Code on Monday, and Cursor already knows it on Tuesday. So does your other laptop. You never explain your project twice.

Most memory starts blank. Yours doesn't have to.

When you install me, you choose: start with the work you've already done with AI, or start fresh. It stays on your computer, and you can delete any of it.

Open source · Self-hosted · Encrypted

I'm Open Agentic Brain and Memory. I live on your machine, and in your private GitHub.

No account, and no server of ours. Free and open source.

Built by Rohith, Chandan and Sushmita

Claude Code Codex Antigravity Cursor Kilo Codebeta Claude Desktop ChatGPT desktop GitHub SQLite

Every new conversation starts with me.

Most AI tools forget everything when the chat window closes. I don't.

Conversation 1 · Mon 10:14

The chat ends. I keep what mattered.

Decisions, fixes and preferences become memories, tagged with the project they belong to.

Conversation 2 · Tue 09:02

A new chat opens. I'm loaded before you type.

A start-up hook hands the model a briefing of this project's recent history: at most 2,000 tokens, ready in 84 ms (p50).

Conversation 2 · Tue 09:20

Need something older? The model asks me.

Over MCP, any connected tool can recall, save and load context on demand. A recall over 10,000 memories takes 28 ms (p50).

Conversation 3 · Wed · other laptop

Even on another machine.

At start-up I quietly merge in what your other machines learned, so the briefing is the same everywhere.

Claude Code · api · conversation 1
You

Let's drop Redis. Sessions live in SQLite.

Claude

Agreed. Removing the Redis client.

You

Also, always use pnpm in this repo.

saved · decision · api saved · preference · api
Claude Code · api · conversation 2
SessionStart · agentic-brain · 84 ms
What happened in api (newest first):
- Tue  bugfix   token refresh race in auth.ts
- Mon  decision sessions live in SQLite, no Redis
- Mon  pref     always use pnpm in this repo
Pinned: deploy runs from main only
1,995 of 2,000 tokens
Cursor · MCP · agentic-brain
→ brain_recall("why did we drop redis", project: "api")
0.92Sessions live in SQLite; Redis only served the old job queue
0.61Deploy health checks: what each one covers
brain_recallbrain_rememberbrain_contextload-my-brain
MacBook Air · ~/code/api
SessionStart
brain merge   pulling from your private repo…
              + 38 memories from MacBook Pro
briefing      same as your other machine

Folders

I file every memory by project.

Each memory is tagged with the repo it came from, found by walking up to the nearest .git. It works the same in every tool.

You open ~/code/web

Open a folder and you get its history first.

The briefing is "what happened in this repo". Its pinned items come first, then pins from anywhere.

Across folders

But nothing is walled off.

A map of up to 20 projects lives in your global CLAUDE.md, AGENTS.md and GEMINI.md. Recall can search one project or all of them. A brand-new repo hears: "Other projects are still searchable."

~/codeexample
api/1,204 memories
web/842 memories
infra/391 memories
new-idea/no memory yet
What happened in web
- Wed  decision  move forms to server actions
- Tue  bugfix    hydration error on /pricing
Pinned (web): staging deploys need a VPN
# ~/.claude/CLAUDE.md · BRAIN:START
Projects: api · web · infra · 17 more
Active this week: web, api
Recurring: auth, deploys, SQLite

Everything you've built, connected.

Every project, and the ideas that connect them. Hover, drag, click.

Example memory with made-up projects. Yours is drawn from your own sessions, in your control panel.

Only you can read me.

I know your decisions, your code and your mistakes. That shouldn't sit on anyone else's server.

Who I belong to

Sealed with your key. Kept in your repo.

I travel as one age-encrypted blob, a Release asset on your private repo. My key lives only in your Keychain. A leaked repo or a leaked token exposes nothing readable.

Why a Release asset and not git? The blob is replaced on every sync. Release assets are replaced in place and don't touch the LFS quota.

But you can read all of me.

Every memory, from every tool and every project, in one place on your own machine.

Open Agentic Brain and Memory · localhostexample
redis
All projectsapiwebinfra
api · decision

Sessions live in SQLite; Redis only served the old job queue

EditPinnedDelete
infra · bugfix

Removed Redis health check from deploy.sh

EditPinDelete
learning · draft

When a cache layer is removed, also remove its health checks and env vars

ApproveDismiss
learning

Run migrations before seeding in CI

Helpful ×6Not helpful

At night, I dream.

By morning I know more than I did, and I've thrown away what didn't matter. Here's the loop I run, every day.

How I work

  1. 01

    Capture

    I read your Claude Code, Codex, Antigravity, Cursor and Kilo Code sessions, first the history already on your machine, then each new one as it happens, into one local SQLite store on your machine.

  2. 02

    Distill

    While you sleep, brain dream turns what I heard into a wiki I keep rewriting. When you finally fix something hard, I draft the lesson for you to approve.

  3. 03

    Rank

    I rank by relevance, then by proof: how often a memory was used, and whether you said it helped. What misleads you sinks.

  4. 04

    Inject

    Before you type, I write a ~2,000-token briefing into CLAUDE.md, AGENTS.md and GEMINI.md, inside my own markers. I never touch your text.

  5. 05

    Measure

    brain eval quizzes me on your own history. If a change makes me remember worse, it doesn't ship.

I check my own work.

On a public long-term memory benchmark, this is how I score.

0end-to-end QA accuracy on LongMemEval-S, 500 questions
0of questions find the answer session in the top 5
0accuracy with 1.27M-token histories (LongMemEval-M)
0vendor servers between you and your memory

LongMemEval-S: Claude reader and judge, k=40 rounds with pseudo-relevance feedback and deep recall (GPT reader: 90.0%). LongMemEval-M: gpt-6-luna reader and judge, same search.

How I compare

Next to the published numbers.

LoCoMo asks questions about long conversations. Here is my score next to giving the same model the whole conversation, and next to the best published memory product.

LoCoMo · answer accuracy (LLM judge, J), categories 1–4
97%of the whole-conversation score, reading 18% of the tokens · +18.2 above the industry best (published)
Whole conversation89.3
Open Agentic Brain and Memory86.6
Industry best*68.4

Open Agentic Brain and Memory and whole conversation: measured by us on the same model (gpt-6-luna answering and judging), all 1,540 questions; with Claude Sonnet answering, Open Agentic Brain and Memory scores 81.8 at 20 memories. *Industry best: the highest score of any memory product in the widely cited 2025 LoCoMo comparison, which used gpt-4o-mini to answer and to judge. Different models, so read that comparison as direction, not a like-for-like race.

LongMemEval-S · answer accuracy by question type, 500 questions
94.2%overall (Claude judge) · 93.4% with a GPT judge · judges agree on 98.8%
Facts you stated100
Things the AI said100
Your preferences100
Updated facts96.2
When things happened91.7
Saying "I don't know"90.0
Across many chats88.7

Claude Sonnet reader, LongMemEval's own judge prompts, k=40 rounds with pseudo-relevance feedback and deep recall. The right evidence was in the model's context for 99.6% of questions. Same search with a GPT reader (gpt-6-luna): 90.0%.

LongMemEval-M · answer accuracy with about 1.27 million tokens of history per question, 500 questions
83.0%with histories too large to hand any model in full · the model read 98.5% fewer tokens · the right evidence reached it for 97.9% of questions
Things the AI said100
Facts you stated94.3
Updated facts92.3
When things happened80.5
Across many chats72.9
Your preferences56.7

The large LongMemEval variant: about 2,500 conversation rounds and 1.27M tokens behind every question. gpt-6-luna answering and judging, deep recall, 40 memories; about 19k tokens read per question. The same setup scores 90.0% on the smaller LongMemEval-S.

How I got better

Every change has to beat the last number.

Nothing about how I search ships on a hunch. Each of these was measured before and after, on the same questions.

Searching by relevance, not recencyrecall on your own history
15% → 65%
Better briefing querybrain eval, 139 questions
69% → 87%
Deeper recallLongMemEval-S answer accuracy
88.2% → 94.2%
Re-ranking the shortlistall evidence in the top 5
85.5% → 94.9%
Following the threadLoCoMo questions that need several chats
65.6 → 80.5
beforeafter0 ──────────── 100

And I'm cheap to remember with.

Better answers from fewer tokens: the model reads what matters, not everything.

LongMemEval-S

75% fewer tokens per question.

About 31.7k tokens of retrieved memory instead of the 127.6k-token history, and 94.2% of answers still right.

LoCoMo

82% fewer tokens per question.

About 3.7k tokens instead of the 20.8k-token conversation, at 86.6% answer accuracy: 97% of what reading the whole conversation scores.

Every session start

A briefing, not a dump.

With 10,000 memories, pasting everything would cost 1.13 million tokens. The context I hand over is about 1,300: 99.88% smaller. The start-up briefing is capped at 2,000, so tools that cut off long start-up messages never drop the end.

Savings meter

I keep score.

Each time a memory saves a tool from redoing work, I log the tokens that work cost, minus what reading the memory cost. Each memory counts once a day, and never below zero.

Whole history127.6k
My memories31.7k

Tokens read per question · same 500 questions

Whole conversation20.8k
My memories3.7k

Tokens read per question · 1,540 questions

Dump all memories1,131,367
My context1,325

Tokens to load context · 10,000-memory store

Control panel · savingsexample
0k tokens saved so far (estimate)
today 18kthis week 96k

Open a new tool tomorrow. I'm already there.

But "supported" can mean three different things. Here they are, kept apart.

Where I live

Every tool you use plugs into me.

"Supported" means three different things. Here they are, kept apart.

ToolCapturePushPull
Claude CodenativeCLAUDE.mdMCP
Codex CLI + appparser + watcherAGENTS.mdMCP
Antigravity CLIhistory + watcherGEMINI.mdMCP
Cursorlocal storeper-repo AGENTS.mdMCP
Kilo Codebetasession parserAGENTS.mdMCP
Claude Desktopserver-sidenoneMCP
ChatGPT desktopby choicenoneMCP

Capture sessions flow back in. Push a digest the tool auto-loads, always present. Pull the model can search memory when it decides to.

A few more things about me

A control panel on localhost.

Search memory, browse the wiki, approve learnings, rate what got injected, and watch capture freshness and sync age.

Open Agentic Brain and Memory control panel demo

Bring your history with you.

Switching from another memory tool? One command imports your whole history, read-only and verified, wires every tool, and can be undone.

Skills that use the memory.

Shipped skills that pull context before work and write decisions back after it.

brain-recallcapture-decisiondebug-systematically onboard-codebasereview-changeswrite-tests

Many machines, one brain.

brain sync on one laptop, brain restore on the next. brain merge reconciles memory from more than one.

It never eats your files.

Every write path is idempotent. Injection lives inside BRAIN:START/END markers, and MCP wiring is additive: it never touches entries it didn't create.

Secrets scrubbed before storage.

Settings are redacted before commit, and memory, wiki and learnings are encrypted. Skills and agents stay plaintext so they stay portable.

I'll be straight with you.

Hosted memory products are good, and easier to set up. Here's the trade, and what I can't do yet.

Hosted memory vs. me

The difference is architecture.

Hosted products solve a real problem well, with zero setup. Here is the trade, including the rows where this side pays.

Hosted serviceOpen Agentic Brain and Memory
Where memory livesVendor's databaseRelease asset on your private repo
Who can read itVendor infrastructureWhoever holds your age key
Key custodyVendor-managedYour OS Keychain, never transmitted
CostPer-seat subscriptionFree
If the vendor shuts downExport, if offeredNothing happens
SyncContinuous, incrementalDaily, one whole encrypted file
SetupSign up, installCLI, a key, a repo
UptimeTheirsNone. It's your machine.

What I can't do yet

Stated plainly.

Say it once. I'll remember it everywhere.

One command wires every tool on your machine, creates your private data repo and generates my key. Then I start with the work you've already done.

Open source · Apache-2.0

I'm awake. Take me home.

Get it on GitHub
curl -fsSL https://raw.githubusercontent.com/rohithhn/Open-Agentic-Brain-and-Memory/main/install.sh | bash
macOS or Linux Node 22+ git A GitHub account