For every AI agent · Self-hosted · EU
Your team forgets overnight. Recallbase does not.
The persistent memory layer for AI-assisted development. Every session recorded, every learning distilled – automatically in your agent’s context on the next start.
$ free plan · no credit card required
Setup happens on your computer. Email yourself the starting point:
Or use it right here: connect Recallbase to the Claude app →
07:58 · Tuesday · your machine
Your agent wakes up – and remembers.
No briefing, no copy-paste, no “where were we?”. The session hook fetches everything that matters.
The problem
Every morning your AI agent starts from zero.
Context windows end, chats disappear, decisions live in people’s heads. That costs the same time every day – and reproduces the same mistakes.
✕Lost context
Last week’s architecture decision? Yesterday’s workaround? Gone the moment the session ends.
✕Repeated mistakes
The agent walks into the same trap as three weeks ago – because nobody wrote down that it is a trap.
✕Knowledge in silos
What project A learned does not help project B. And the team wiki has been outdated since the day it was written.
Who it's for
Not just for developers.
Recallbase gives any AI agent a memory – anywhere teams work with AI on knowledge over weeks.
Software teams
Architecture decisions, workarounds and yesterday's state – remembered across projects, for Claude Code, Codex, Gemini & Cursor.
Newsrooms
Research, sources and the thread of a story persist – even when several people work on it.
Firms & consulting
Case context, deadlines and earlier assessments stay retrievable – GDPR-compliant in the EU or self-hosted.
Research & knowledge work
Literature, notes and confirmed insights grow into a searchable memory instead of sinking into chats.
Features
A memory that writes itself.
Recallbase works in the background of your sessions – no upkeep, no copy-paste, no “documentation Friday”.
Context from second one
On session start, Recallbase injects project history, recent decisions and confirmed knowledge straight into your agent’s context.
Session hookLearning loop
Observations from your sessions are collected, condensed and only become permanent knowledge after your confirmation. No noise, no drift.
Observe → ConfirmKnowledge graph
Insights, patterns and people linked across projects. What one project learns, all of them know.
Cross-projectTenant isolation
Row Level Security separates organizations directly in PostgreSQL – fail-closed. A mistake returns nothing instead of everything.
Postgres RLSSearch that understands questions
“Where was that printer issue?” finds the Cura session – even with zero matching words. Semantic search across all sessions and your entire knowledge.
Semantic + full textYour memory dreams at night
Crashed or undistilled sessions are automatically post-processed overnight. By morning the memory is complete – without anyone having to think about it.
Auto-dreamingContradiction detection
New knowledge automatically retires outdated facts – invalidated, not deleted. Your agent never re-learns old mistakes, and "what was true back then?" stays queryable.
Bi-temporalPrivate stays private
Anything between <private> and </private> never reaches the database. Sensitive details stay in the session – guaranteed.
Your memory, visible
"My memory" in your account: browse every session, search everything – you can always see what your agents know.
Memory viewerReminds you before you ask
Save an intention with a trigger – "when working on the auth module again, fix the token bug first". Recallbase surfaces it actively at the right moment, instead of waiting for you to search.
Prospective memoryLearns rules for the next project
What keeps recurring across sessions, Recallbase distills overnight into cross-project rules – "deploys here usually fail on missing ENV vars". Knowledge moves from the silo to the organization.
Schema inductionA working profile you can audit
From actual behavior emerges an evidence-backed profile: working style, strengths, recurring mistakes – every claim tied to a source session. Not an invented persona block, but traceable and correctable.
Auditable self-modelFalse knowledge gets muted
Outdated or refuted facts aren't just flagged – they're actively kept out of context, weighted by origin and trust (your statement beats a web find). And you can trace forensically what a poisoned entry influenced.
Trust & suppressionWhat you use stays on top
Every memory carries a stability that grows with each real retrieval – the forgetting curve from spaced-repetition research. Rarely used knowledge drifts down the ranking instead of silting up the index. Nothing is deleted: everything stays findable.
Forgetting curveWhat surprised it is forgotten last
On write, Recallbase measures the distance to the most similar thing already there. What was new – and especially what refutes existing knowledge – gets more weight in ranking and is distilled first at night. Like the brain, where replay follows the reward-prediction-error.
Surprise scoreDashboard & team management
API keys, invites, subscription – all in your browser account. New members join via link and instantly see your organization's projects.
recallbase.de/accountAPI-first & self-hosted
Claude hooks, an MCP server for Codex/Gemini/OpenCode/Cursor, and a clean REST API. One memory, every agent.
Your infrastructureHow it works
Three steps that happen on the side.
Session starts
The hook calls Recallbase, fetches project context, recent sessions and master memory – and your agent knows where you left off yesterday.
While you work
Recallbase watches signals in the transcript: corrections, confirmations, patterns. They become evidence-backed observations – automatically.
Session ends
The session is summarized and stored in a structured way. Confirmed insights move into the knowledge graph – ready for session n+1.
Performance
An agent with a memory is a better agent.
A memory doesn't just save typing – it makes the agent more effective. The reason isn't a trick, it's mechanics: a small, curated context instead of a stuffed instruction file.
Median, measured across 1,663 real sessions over 127 projects in our own production use.
Fewer tokens
A curated mini-context instead of full history – cheaper and faster per answer.
Higher rule-following
Models follow short contexts more reliably. No rule-noise from 800 lines.
No cold start
The agent knows history and decisions from line one – fewer false starts.
No repetition
Confirmed knowledge keeps the agent from re-learning solved mistakes.
In a controlled evaluation (12 project-specific questions), an agent without memory answered only 1 of 12 correctly – with Recallbase's curated context 12 of 12, as well as with the full history but at fewer tokens. Mechanics, not a marketing number. Why long CLAUDE.md files tip over →
Security
Isolation is not a convention here. It is database law.
- Row Level Security on every table – tenants only see their own data, enforced by PostgreSQL itself.
- Fail-closed – without a set identity, every query returns zero rows. A programming mistake never becomes a data leak.
- Your infrastructure – self-hosted on your server or as an EU instance. TLS, API keys, full control.
Privacy & Compliance in detail – DPA, EU residency, your rights →
Pricing
One memory that grows with you.
Start free. No setup fee, cancel anytime. Prices incl. VAT.
Free
0 € /month
- 1 project
- Full memory loop
- Session history & knowledge graph
- All agents (Claude Code, Codex, Gemini, MCP)
Pro
19 € /month
- For individuals
- Unlimited projects
- Semantic search & knowledge graph
- Priority support
Team
39 € /seat · month
- Everything in Pro
- Shared team memory
- Invites & roles
- Cross-project insights
Enterprise
Let's talk
- Self-hosted in your infrastructure
- SSO & role model
- DPA, EU data residency
- Support SLA & onboarding
Pay yearly and get 10 months instead of 12. All plans self-hosted or as an EU instance. Questions? Write us.
Comparisons
Why not just …?
… a CLAUDE.md?
Because it grows until the agent ignores rules – and you pay full price every session.
Read comparison →… Mem0?
Mem0 is an SDK you build with. Recallbase is the finished service – no code.
Read comparison →… claude-mem?
Great for solo + one machine. Recallbase is the service for teams, device sync and EU compliance.
Read the comparison →… a wiki or Notion?
Because a wiki has to be maintained. Recallbase writes itself.
Read comparison →Try it live
Ask a stranger's memory.
This is the real product running against a demo project ("Webshop Aurora", 7 stored sessions). Ask a question the way you would ask your agent:
Not claimed – formed
What a mature memory knows on its own.
Recallbase distilled this itself from the demo project – cross-project rules and an evidence-backed working profile. No manual entry, no claim: this is the system's output.
New: on your phone too
Your memory in your pocket.
📱 Recallbase in the Claude app
Connect your project memory to the Claude app or ChatGPT as a connector: check status, search knowledge, save notes – on the go, no installation. Phone and computer share the same brain.
Connect in 60 seconds →How it works
An MCP connector instead of another app: why your AI memory no longer has to stay at your desk – the article.
Read the article →Trust
Hardened in daily use.
of daily production use before launch – not a prototype, a proven tool.
documented sessions across 127 of our own projects before we opened it up.
Tenant isolation as a database law. Self-hosted on your server or as an EU instance.
Recallbase just opened up – the first user voices will appear here. Be one of the first →
FAQ
Frequently asked questions
Claude now has its own memory – why Recallbase?
Claude's native memory is a personal profile (role, preferences, style) inside Anthropic. Recallbase is project-scoped working memory across all agents: Claude Code, Cursor, Codex, Gemini and the phone app share the same state – including team sharing, EU hosting and export. They complement each other; neither replaces the other.
What is Recallbase?
Recallbase is a persistent memory layer for AI-assisted development. It records every session, distills insights from it and automatically loads project history and confirmed knowledge into your AI agent’s context on the next session start.
Does Recallbase work with Claude Code, Codex, Gemini & co.?
Yes. Claude Code gets session hooks (automatic, crash-safe). Every other agent – Codex CLI, Gemini CLI, OpenCode, Cursor, Windsurf – connects via the built-in MCP server. And the REST API is open to any tool that speaks HTTP.
Where is my data stored?
On your own server or in an EU instance. Recallbase is self-hosted-first: one PostgreSQL database, one service, your infrastructure. Tenant isolation is enforced with Row Level Security directly in the database.
How is Recallbase different from a wiki?
A wiki has to be maintained – Recallbase writes itself. Summaries come from real sessions, observations become confirmed insights, and everything is injected exactly where it is needed: into the context of the next session.
Why not just a CLAUDE.md?
Because long instruction files tip over: past a few hundred lines the agent starts ignoring rules – while every session still pays the full token price. Recallbase keeps context small and curated: confirmed insights instead of instruction sprawl, project-relevant knowledge hard-capped, details on demand via search. And it distributes across your team automatically – no copy-pasting into everyone's CLAUDE.md.
What does Recallbase cost?
Free stays free (1 project, individual). Pro is €19 per month for individuals with unlimited projects. Team is €39 per seat per month and unlocks shared team memory, invites and roles. Pay yearly and get 10 months instead of 12. Self-hosted and larger organizations: Enterprise on request.
Give your team a memory.
Set up in two minutes. Start free, no credit card required – upgrade when your memory grows.
Start for free