AlekseiMarchenko/central-intelligence

**Agents forget. CI remembers.**
Persistent memory for AI agents. Store, recall, and share information across sessions. Works with Claude Code, Cursor, LangChain, CrewAI, and any agent that supports MCP.
**CI never rewrites your memories.** Facts are extracted for search, but your content is always returned verbatim. No junk memories, no hallucinated rewrites, no data loss.
-52.2%25-6d5aff?style=for-the-badge)
-22c55e?style=for-the-badge)
Quick Start (30 seconds)
Or run locally with no cloud:
When to Use Central Intelligence
> **Heuristic:** If you would write it in a note to your future self, store it in Central Intelligence.
| Scenario | What to do |
|----------|-----------|
| Starting a new session, need context from before | recall or context |
| Discovered something important (architecture, preferences, fixes) | remember |
| Multiple agents working on the same project | share with user/org scope |
| You keep re-learning the same things each session | remember once, recall forever |
| Handing off a task to another agent or session | remember key decisions, next agent calls context |
| User tells you the same preferences repeatedly | remember them, check with recall next time |
**Don't store:** secrets, passwords, API keys, PII, large binary files, or ephemeral scratch data.
The Problem
Every AI agent session starts from zero. Your agent learns your preferences, understands your codebase, figures out your architecture — then the session ends and it forgets everything. Next session? Same questions. Same mistakes. Same context-building from scratch.
Central Intelligence fixes this.
What It Does
Five MCP tools give your agent a long-term memory:
| Tool | Description | Example |
|------|-------------|---------|
| **remember** | Store information for later | "User prefers TypeScript and deploys to Fly.io" |
| **recall** | Semantic search across past memories | "What does the user prefer?" |
| **context** | Auto-load relevant memories for the current task | "Working on the auth system refactor" |
| **forget** | Delete outdated or incorrect memories | forget("memory_abc123") |
| **share** | Make memories available to other agents | scope: "agent" → "org" |
Benchmarks
LifeBench (2026) — Long-Term Multi-Source Memory
CI scores **52.2%** on LifeBench, the hardest published memory benchmark (2,003 questions across 10 users, 51K real-world events including messages, calendar, health records, notes, and calls).
| Overall | Info Extraction | Multi-hop | Temporal | Nondeclarative |
|---------|-----------------|-----------|----------|----------------|
| **52.2%** | **47.2%** | **52.9%** | **46.4%** | **64.1%** |
Answer model: gpt-5.4-mini. Judge: gpt-4.1-mini. Evaluation harness: lifebench-eval.
LongMemEval (ICLR 2025) — Conversational Memory
CI scores **75.0%** on LongMemEval, testing conversational memory across 500 questions spanning single-session recall, multi-session reasoning, temporal reasoning, knowledge updates, and preference tracking.
| Overall | Single-session | Multi-session | Temporal | Preference |
|---------|----------------|---------------|----------|------------|
| **75.0%** | **91.9%** | **66.2%** | **69.9%** | **76.7%** |
Answer model: gpt-5.4-mini. Judge: gpt-4o. Evaluation harness: lifebench-eval.