# mismem [Health: Active]

**Category:** 🗄️ Databases  
**Repository:** https://github.com/lucasmella-stack/misMEM  
**GitHub Stars:** 3  
**npm Downloads (last month):** 119  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/mismem

## Description
Cooperative persistent memory for LLMs: episodes, memories and traits over SQLite (FTS5, Hebbian)

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

```json
"mcpServers": {
  "mismem": {
    "command": "npx",
    "args": ["-y","@lucasmella/mismem"]
  }
}
```

## Documentation

## What mismem MCP server does

The mismem MCP server provides a shared memory layer for multiple MCP clients and agent instances. Instead of keeping context inside one assistant or application, it stores data in a SQLite database that can be accessed by local clients through the same MCP endpoint. The default database path is `~/.mismem/mem.db`, and the location can be changed with `MISMEM_DB`.

Its data model has three levels. Episodes retain dated source text for immediate context, but are intended to expire after consolidation. Memories represent distilled knowledge and have a salience score that decreases over time unless recalled. Traits store stable preferences, decisions, or rules and remain until explicitly removed. This model is useful when raw conversation history should not persist indefinitely while conclusions and durable policies should remain available.

## How it works

The mismem MCP server exposes five primary operations:

- `capture` records an episode under a hierarchical scope.
- `recall` searches stored information and reinforces returned memories.
- `consolidate` turns related episodes into a memory.
- `crystallize` records a trait supported by evidence.
- `forget` removes memories that meet the configured forgetting conditions.

Search uses SQLite FTS5. Memory recall can also add semantic matching through a local Ollama embedding model, with vectors stored in the same database. Without Ollama, recall falls back to FTS5. Semantic search applies to memories; episodes and traits use FTS5. Salience decays to half its value after 90 days by default, and recall applies the pending decay before adding Hebbian reinforcement.

A compatibility layer also supports Engram-style operations such as `mem_save`, `mem_search`, `mem_context`, and session-related calls. Hierarchical scopes are used to separate contexts and prevent cross-scope leakage. Capture and ingestion include deduplication so repeated imports are idempotent.

## Setup and configuration

The quickest installation uses the npm package `@lucasmella/mismem` with `npx` in an MCP client's server configuration:

```json
{
  "mcpServers": {
    "mismem": {
      "command": "npx",
      "args": ["-y", "@lucasmella/mismem"]
    }
  }
}
```

The local engine requires Node.js 22 or newer. Developers can instead clone the repository, install dependencies from `engine`, build the TypeScript project, and launch `engine/dist/server.js` with Node. The database is created on first use.

For remote access, the project provides a Docker image with Streamable HTTP support. Mount a volume at `/data`, publish port 3200, and set `MISMEM_AUTH_USER` and `MISMEM_AUTH_PASS` when authentication is needed. The HTTP service also exposes `/healthz`; the included viewer provides read-only pages and APIs for statistics, recent entries, and search. Do not expose HTTP without authentication beyond localhost.

## Tools and capabilities

The mismem MCP server includes command-line utilities for ingesting ChatGPT exports and VS Code transcripts, consolidating data, forgetting entries, viewing statistics, migrating Engram data, and managing embeddings. A scheduled LLM consolidation workflow can use OpenRouter to distill episodes into memories, with a Pareto pre-filter intended to remove noise. Optional embedding backfill and reindexing are available through `mismem-embed`.

## Limitations and notes

SQLite uses WAL mode, supporting multiple readers but one concurrent writer. Semantic search requires a running local Ollama service and the `nomic-embed-text` model; otherwise the system remains usable with lexical search. Traits and episodes do not receive semantic search. Memory deletion is threshold- and grace-period-based, and `forget` supports a dry-run mode before changing the database.

The project is marked as having a Bun-compiled distributable still in progress. Personal databases, exports, and dumps should stay out of version control. The software is licensed under MIT.

_Full upstream README: https://allmcps.com/mcp/mismem/readme_