CogniCore MCP
Memory for agents: a Claude Code session resumes in Codex with the same decisions and open tasks.
- 83 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- Hooks read session transcripts and write a summary to disk; with sync enabled the redacted brief goes to a hosted server.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add mcp/cognicoreDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Hooks read session transcripts and write a summary to disk; with sync enabled the redacted brief goes to a hosted server.
This is third-party code. Review the repository files before installing.
What it does
CogniCore stores the outcome of a work session and feeds it into the next. Hooks read the agent transcript at session end and extract the goal, current state, open tasks, decisions with reasons and exact values such as ticket IDs, versions and paths. At the next start the agent receives a brief of roughly 450 tokens. Memory stays local in SQLite under `~/.cognicore/mcp` and secrets are redacted before saving. The author compared token use with claude-mem on real Claude Code and Codex sessions. An optional `cognicore login` syncs with claude.ai, Claude Desktop and mobile through a hosted MCP server with 31 tools. Besides the plugin, the package ships a Python library for memory, reflection and safety checks with no vector database.
Who it is for. People who work across several agents and are tired of re-explaining project context in every new session.
Good fit when
- A task started in Claude Code must continue in Codex.
- A new session should know yesterday's decisions and the open next step.
- An agent in claude.ai reads the same memory after signing in.
- A brief is wanted instead of a full chat log to save tokens.
Not a fit when
- Memory cannot leave the machine even in redacted form, yet sync is enabled.
- Tasks are short and never repeat, so cross-session context adds nothing.
- You need shared team memory with roles and permissions.
Example request
Load the saved project brief and continue from the open step: which decisions were made and what is left.MCP
- Transport
- stdio, http
- Authentication
- OAuth
Limitations
Install with `pipx install "cognicore-env[mcp]"` and the `cognicore install` command. In Codex, open `/hooks` once and trust the three CogniCore hooks. The hosted server requires GitHub sign-in and gives each account its own database; only the redacted handoff leaves the machine, never the transcript. Secrets are redacted automatically, yet private material is worth checking in the local database. The claude-mem comparison was run by the author, so reproduce it with `scripts/bench_realworld.py`. The Python library needs Python 3.10 or newer.
How to disable. Remove the CogniCore plugin through `/plugin` and delete its hooks from the Claude Code and Codex configuration, then erase `~/.cognicore`. Close a hosted account separately.
FAQ
How do I install CogniCore for Claude Code and Codex?
With `pipx install "cognicore-env[mcp]"` and `cognicore install`. Or inside Claude Code: `/plugin marketplace add cognicore-dev/cognicore-env` then `/plugin install cognicore@cognicore-beta-marketplace`.
Where is memory stored?
Locally in SQLite at `~/.cognicore/mcp`. `python -m cognicore.memory.viewer` opens a read-only page listing every saved session.
Which MCP tools does the agent call?
`cognicore_save_session` and `cognicore_resume_session` carry work across. The hosted server has 31 tools in total, listed in `docs/mcp-tools.md` with notes on whether each reads or changes data.
How do I connect the hosted server?
Add `https://mcp.cognicoree.tech/mcp/stream` as a custom connector and sign in with GitHub. The `/health` path on the same domain reports its state.
Does it need a vector database or an API key?
No. Memory, reflection and safety checks need no key, and retrieval runs in-process with BM25 and multi-hop search.
Discussion
No questions or reviews yet.
Related
MCP serversCortex keeps decisions and project context between Claude Code and Codex sessions in a local database and shows what it recalled
Knowledge-graph memory for agents: Claude Code and Codex plugins, an MCP server and a Python library
The session-recall command reads an agent’s local sessions and returns past conversations in compact form, with no MCP or hooks.
Shared memory for coding agents: messages in SQLite, old context compacted into a summary tree and MCP tools to search it.
Coding memory across sessions, machines and teammates: decisions and dead ends flow into later prompts on their own.
hipocampus gives an agent file-based memory: a 3,000-token topic index loads each session and shows what came up before
README in short
The project combines a plugin that hands sessions between agents with a Python memory library. The README covers installation, hook behavior, the hosted server and token-use comparison results.