Memory Backend — CrewAI Integration

CrewAI agents never forget anything.

Mimir plugs into CrewAI (54K GitHub stars) as a persistent cross-session memory backend. Built by community contributor Jacopos311 — 13 commits, 28 reviews, mergeable. One pip install. Agents that remember across sessions, crews, and deployments.

CrewAI Stars

54K

CrewAI is the leading multi-agent orchestration framework. 54K GitHub stars, production deployments.

PR Status

13/28

13 commits, 28 reviews on PR #6208. Mergeable. Community-owned from day one.

Setup

1 install

pip install mimir. Add one config block. Every crew gets persistent memory.

How It Works

Step 1

Feature Request

Filed #6168 on June 15 — requested Mimir as a CrewAI memory backend option.

Step 2

Community Builds It

Jacopos311 shipped PR #6208 — Mimir toolset via subprocess, MCP-native. 13 commits.

Step 3

Crews Remember

Every CrewAI agent gets persistent memory. FTS5 recall, confidence decay, cross-session compounding.

The Architecture

CrewAI Agent
  └─ Memory Backend (abstract)
      └─ MimirProvider
          ├─ mimir CLI (subprocess)
          ├─ MCP stdio transport
          ├─ SQLite + FTS5
          └─ 23 MCP tools (recall, store, decay, link, journal…)

Quickstart

install

# Install Mimir (the memory engine)
curl -sSL https://raw.githubusercontent.com/\
Perseus-Computing-LLC/mimir/main/\
scripts/bootstrap.sh | bash

# Or pip
pip install mimir

configure

# crewai.yaml — add Mimir as memory backend
memory:
  provider: mimir
  config:
    db_path: ~/.mimir/data/mimir.db

# Agents now persist across sessions

What It Unlocks

Cross-Session Memory

Crew runs a task today. Tomorrow's crew recalls yesterday's decisions. No re-orientation. No token tax.

Cross-Crew Knowledge

Crew A learns a pattern. Crew B sees it. FTS5 keyword search surfaces relevant history across all crews.

Confidence Decay

Un-reinforced facts fade. Reinforced facts stick. Agents know what they know — and what they're uncertain about.

Zero Cloud Dependencies

SQLite + FTS5. No API keys, no network calls, no embeddings model. Runs on the same machine as your crew.

PR #6208 on CrewAI → Mimir on GitHub →