Central Intelligence (CI for short) is a persistent memory tool for AI agents, open-sourced on GitHub. Its core idea is: every session of an AI agent starts from scratch and forgets previously learned user preferences, codebase understanding, and architecture information; CI solves this problem by storing, retrieving, and sharing information across sessions.
Main Capabilities
CI provides five MCP tools that give agents long-term memory. It claims it never rewrites the user's stored memories: content is extracted into structured facts for search, but the original content is always returned verbatim, thereby avoiding "junk memory," hallucinated rewrites, and data loss.
Memories are broken down into facts containing entities, temporal information, and causal relationships. Retrieval uses a dual-path architecture: fact-based four-way search (vector, BM25, graph traversal, temporal) runs in parallel with memory-based two-way search. A query type classifier routes each question to the best retrieval path, results are fused via Reciprocal Rank Fusion, and reranked using a local cross-encoder model.
CILocal can read the configuration files of 5 AI coding platforms, making them searchable alongside stored memories. Memories stored through Claude Code can also be discovered in Cursor, and vice versa. Retrieval responses include source (which tool the memory came from), freshness score (freshness_score), and duplicate grouping (duplicate_group, used for approximate deduplication across tools).
How to Use
Quick start can be completed in 30 seconds, and local running without cloud services is also supported. Ways to connect to MCP clients include:
- Claude Code: add configuration under mcpServers in ~/.claude/settings.json.
- Cursor: add configuration in ~/.cursor/mcp.json.
- Any MCP-compatible client: the MCP server is published on npm as central-intelligence-mcp; just point to it and set the CI_API_KEY environment variable.
The cloud API base address is https://central-intelligence-api.fly.dev，所有端点需要 Authorization: Bearer <api-key> request header. Available endpoints include:
- POST /memories/remember
- POST /memories/recall
- POST /memories/context
- DELETE /memories/:id
- POST /memories/:id/share
In addition, it can also return the memory count, usage events, and active agent information for an authenticated API key. Deployment to Fly.io is also supported, after which you can point the MCP server to your own instance.
Performance Evaluation
The project has published three benchmark results:
- LifeBench (2026) — long-term multi-source memory: CI scored 52.2%; the benchmark contains 2,003 questions from 10 users and 51K real-world events (messages, calendars, health records, notes, calls). The answering model is gpt-5.4-mini, the judging model is gpt-4.1-mini, and the evaluation tool is lifebench-eval.
- LongMemEval (ICLR 2025) — conversational memory: CI scored 75.0%, testing 500 questions covering single-session recall, multi-session reasoning, temporal reasoning, knowledge updates, and preference tracking. The answering model is gpt-5.4-mini, and the judging model is gpt-4o.
- Agent Memory Benchmark (AMB) — infrastructure testing: uses the open-source Agent Memory Benchmark to compare with other providers. The project notes that AMB is maintained by the same author as Central Intelligence, recommends running it yourself for verification, and welcomes PRs for new provider adapters.
Applicable Scenarios and Limitations
Applicable heuristic rule: if you would write it in a note to your future self, store it in Central Intelligence.
Should not store: secrets, passwords, API keys, personally identifiable information (PII), large binary files, or temporary draft data.
Advanced retrieval features — fact extraction, entity graphs, multi-hop reasoning, temporal inference, explainability tracing — have been prototyped in the codebase and are planned to be offered in the enterprise edition. For architecture details, see the v1.0.0 prototype release notes; for commercial availability, see the pricing page (centralintelligence.online/#pricing).
Contributions to the project are welcome; you can submit an issue or a PR.
