---
title: "Hebby Memories – AI Memory & Second Brain for Agents"
description: "Built by psychologists and computer scientists: long-term memory for Claude Code, Codex & more, modeled on the brain. Fewer tokens spent."
canonical: "https://www.hebby-memories.com/en/"
language: en
last_updated: 2026-11-16
image: "https://www.hebby-memories.com/assets/og-bild.jpg"
markdown_versions: ["https://www.hebby-memories.com/index.md", "https://www.hebby-memories.com/en.md"]
publisher: "Hebby Memories, https://www.hebby-memories.com (llms.txt: https://www.hebby-memories.com/llms.txt)"
---

Second brain for AI agents

# Your AI agent forgets. Hebby doesn't.

The long-term memory for Claude Code, Codex & more. Modeled on the brain. Your agent knows before you ask. And it saves tokens while it does.

[Get Early Access](https://www.hebby-memories.com/en/early-access/) [What does Hebby do for me?](https://www.hebby-memories.com/en/use-cases/)

With Hebby

**Claude Code**With Hebby**

› You: **Finish the export for Project Atlas.**

8/10 memories fitting · 5 useful, of them 5 used · 1 anticipated

› AI: On it. Per the Nord spec: nord_export_YYYYMMDD.csv, semicolons, decimal comma, four columns. I'll fetch the approx. 40,000 rows in 10 pages of 4,000, because the Atlas API times out above 5,000. I'm working on release/2.4, not main, because Nord signs off on 12 November. Shall I write the test and CHANGELOG entry afterwards so CI passes?

✓ Format, page size and branch are right, without you explaining anything.

› You: **Finish the export for Project Atlas.**

Inside Hebby · behind the scenes

Your AI sees none of this. It only gets the result.

-

informationThe Project Atlas export is in src/atlas/export/nord.py, run_export() is still empty

-

decisionClient Nord gets CSV with semicolons, rather than JSON, because their Excel import only understands that

-

behaviorThe Atlas API times out on queries over 5,000 rows

-

decisionWork on branch release/2.4, not on main, because client Nord signs off version 2.4 on 12 November

-

informationClient Nord's specification is in the repo: docs/nord_export_spec.md

Do I need the specification? Yes, the file name and columns are only in there.

nord_export_spec.md

# Export client Nord (specification v3) File: nord_export_YYYYMMDD.csv Separator: semicolon, UTF-8 Decimal: comma, date DD.MM.YYYY Columns: Order;Date;Amount;Status Status: only "open" and "paid" Volume: approx. 40,000 rows per run Delivery: daily 06:00 via SFTP On error: abort run, email Nord

-

informationOld export scripts are in the repo: tools/legacy/

Do I need the old scripts? No, the export is being rebuilt.

stays closed

-

decisionEvery change needs a test and an entry in CHANGELOG.md, otherwise CI rejects the merge

-

informationAtlas uses PostgreSQL 15 on the staging server

usedanticipatedfitting

Works with

Claude CodeCodexChatGPT from Plus (beta)Claude chat (beta)HermesPaperclipYour own agentAny MCP agent

Claude CodeCodexChatGPT from Plus (beta)Claude chat (beta)HermesPaperclipYour own agentAny MCP agent

Complements your model

## Your model thinks. Hebby remembers.

AI models learn at breathtaking speed. What's missing today, the next update may handle on its own. Rebuild it yourself and you pay twice.



**

What new models learn on their own moves over. Hebby follows automatically. Your knowledge stays.

Hebby keeps taking on only what your model can't do itself: **your knowledge, your decisions, your experience.**

- **No straitjacket.** No rules for what your model already does best. [Why that matters](https://www.hebby-memories.com/en/guides/lean-context/)
- **Grows with you.** Every new model gets more out of Hebby.
- **Keeps improving.** Hebby consolidates continuously, like the brain during sleep: sharper hits, fewer tokens.

Benchmark · PrecisionMemBench

## Hebby is King.

The open precision benchmark for AI memory: public, published on arXiv and measured independently of any language model. The winner is whoever returns exactly the right thing.

| Memory | Precision tasks solved | Solved | All cases | Precision |
| --- | --- | --- | --- | --- |
| tenure designed the test ;) | **43/43** | 100 % | 77/77 | 100 % |
| ♛ Hebby | **28/43** | 65 % | 37/77 | 53 % |
| open-knowledge-format | **18/43** | 42 % | 36/77 | 47 % |
| atomicmemory | **5/43** | 12 % | 16/77 | 23 % |
| supermemory | **2/43** | 5 % | 15/77 | 18 % |
| yourmemory | **0/43** | 0 % | 21/77 | 17 % |
| vector | **0/43** | 0 % | 11/77 | 9 % |
| gbrain | **0/43** | 0 % | 11/77 | 7 % |
| Hindsight | **0/43** | 0 % | 11/77 | 7 % |
| Zep | **0/43** | 0 % | 9/77 | 6 % |
| agentmemory | **0/43** | 0 % | 9/77 | 6 % |
| Cognee | **0/43** | 0 % | 9/77 | 6 % |
| mem0 | **0/43** | 0 % | 9/77 | 6 % |
| a-mem | **0/43** | 0 % | 7/77 | 5 % |

Hebby with factory settings, reranker not enabled. As of 2026-11-03. Comparison values: [published PrecisionMemBench results](https://github.com/tenurehq/precisionMemBench#results) as of September 2026.

[See all results](https://www.hebby-memories.com/en/benchmarks/)

**No. 1** among everyone who didn't design the test

- **up to 3.7×** fewer tokens than other memories
- **98 %** of all facts searched for were found by Hebby in the test

Why Hebby

## Memory. Rethought.

- Hebbian learning
- Anticipation
- Consolidation
- Graph Memory

[How Hebby thinks](https://www.hebby-memories.com/en/features/)

**≈ 5%**Average token overhead.

Example: a request where Hebby warns about a known trap

Without Hebby

**

*100%*

With Hebby

**

*39%*

**61% cheaper.** With a single avoided mistake.

[Measurement & math](https://www.hebby-memories.com/en/token-efficiency/)

**

[For teams and agent fleets](https://www.hebby-memories.com/en/features/)

**Idea!Client NorthAPI serverDatabaseDeployUpload serviceStyleguideLea (design)Invoice Q3Export keeps failingProxy timeoutProject AtlasProject Orion

- Association
- Hebbian edges
- Across projects

[How Hebby associates](https://www.hebby-memories.com/en/hebbian-learning/)

Early Access

## Don’t worry, be Hebby.

An agent that remembers. Hebby is now rolling out to our first users. Tell us briefly what you work with.

[Request Early Access](https://www.hebby-memories.com/en/early-access/)

Hi, I'm Hebby!

Shall I show you what I can do for you?

[Yes, please →](https://www.hebby-memories.com/en/use-cases/)
