# cognicore-env

**Category:** 🧠 Knowledge & Memory  
**Repository:** https://github.com/cognicore-dev/cognicore-env  
**GitHub Stars:** 77  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/cognicore-env

## Description
Persistent memory, reflection & safety for AI agents. Zero dependencies — no vector DB, no embedding server. BM25 + multi-hop graph retrieval. MCP server with 10 tools.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "cognicore-env": {
    "command": "uvx",
    "args": ["cognicore-env"]
  }
}
```

## Documentation

## What cognicore-env does

The cognicore-env MCP server adds a persistent memory and reflection layer for AI agents. It preserves more than conversation text: agents can retain experiences, failed approaches, successful outcomes, and supporting context across sessions and processes. Reflection can derive lessons from repeated failures and make those lessons available before a later attempt.

The project also includes safety checks intended to block prompt injection and jailbreak attempts before they reach an agent. Decision events can be recorded for later replay, and a run can be branched from a selected step with modified settings. Memory can be exported as a signed bundle and imported with verification; unverified records are placed in quarantine rather than normal search.

This places cognicore-env in the 🧠 Knowledge & Memory category. It is a fit for agents that need continuity between runs, reviewable execution history, or verified transfer of learned memory.

## How it works

The base runtime operates in-process on Python and does not require a vector database, embedding server, or API key for memory, reflection, or safety. Retrieval starts with lexical matching and can use multi-hop graph traversal. The multi-hop process identifies query entities, finds anchor records, follows session and time relationships, and selects related records that collectively cover the query.

The MCP integration exposes ten tools to MCP clients, allowing the client to use CogniCore capabilities without adding application-specific integration code. A separate hosted endpoint is available for clients that support custom MCP connectors. That service uses GitHub sign-in and assigns each GitHub account its own memory database.

## Setup and configuration

Install the package with:

```bash
pip install cognicore-env
```

Python 3.10 or newer is required. The local MCP configuration shown by the project runs through `uvx`:

```json
{
  "mcpServers": {
    "cognicore-memory": {
      "command": "uvx",
      "args": ["cognicore-env[extension]", "cognicore-extension"]
    }
  }
}
```

The hosted MCP endpoint is `https://mcp.cognicoree.tech/mcp/stream`. Its health endpoint is `https://mcp.cognicoree.tech/health`. The base package has no mandatory dependencies; optional extras add semantic memory, LLM agents, a dashboard server, reinforcement-learning support, mem0 transfer, or development tooling.

## Tools and capabilities

The cognicore-env MCP server supports these documented capability areas:

- Persistent memory storage and retrieval across sessions and processes.
- Reflection over repeated failures and lessons learned.
- Prompt-injection and jailbreak safety classification.
- Event recording, replay, and branching from a prior step.
- Signed memory export and fail-closed import with quarantine for unverified data.
- Lexical retrieval with multi-hop session and time-link traversal.

Memory can optionally use semantic search through the `memory` extra. The package also provides command-line functions for listing environments, training, benchmarking memory against no-memory runs, running agent tournaments, and opening its dashboards or observability studio.

## Limitations and notes

Lexical retrieval can miss relevant records when the wording used in a session differs substantially from the query. The project documents semantic search as an optional feature rather than part of the zero-dependency base installation. Multi-hop retrieval improves results when several sessions must be combined, but its reported advantage decreases when a much larger context window already contains enough records.

The hosted service is separate from the local process configuration and requires GitHub sign-in. The README names Claude, Claude Code, Cursor, and Sarvam Cowork for the hosted connector, while the local MCP setup is intended for MCP-compatible clients generally. No individual MCP tool names are provided in the supplied material, so tool-level calls cannot be specified here.

_Full upstream README: https://allmcps.com/mcp/cognicore-env/readme_