Context
Agent memory often relies on opaque vector search with no ranked recall metrics, no token discipline, and no visibility into what was excluded or why context was injected.
Projects
Mimir
Mimir is self-hostable agent memory with episodic, semantic, and procedural layers, a six-provider retrieval orchestrator, and MCP plus REST APIs for Cursor, Claude Code, Codex, and automation. Status: Live / Released (Apache-2.0, v0.1.0-rc4). Each recall returns token-budgeted context and a full trace.
Live / Released
AI Infrastructure
Self-hosted long-term memory for AI agents
Approach
Mimir fuses six retrieval providers with trust scores and lifecycle rules, enforces token budgets with exclusions logged in the trace, and exposes dashboards, benchmarks, and MCP tools for remember/recall/outcome workflows.
Features
Episodic, semantic, and procedural memory layers
Six-provider retrieval orchestrator with task-aware weighting
MCP Streamable HTTP and REST slash-command API
Dashboard: benchmarks, token efficiency, world ranking tab
memory_recall / memory_remember / memory_record_outcome loop
Quarantine, trust scores, and memory lifecycle
Daily benchmark job and rerunnable retrieval harness
Cursor, Claude Code, and Codex integration packs
Capabilities
- Agent-memory architecture with episodic, semantic, and procedural layers
- Six-provider retrieval orchestration
- MCP and Cursor-native agent integrations
- Benchmark harness and published ranking methodology
- Token-budgeted context with trace logging
- Self-hosted Docker deployment (SQLite or Postgres)
Technologies
PythonFastAPIMCPSQLAlchemyChromaDBDockersentence-transformersAPScheduler
Notes
- Open-source stack the lab uses in production agent workflows.
- Recall traces show what was retrieved and excluded instead of black-box RAG.
- Token budgets and benchmarks keep memory evaluation rerunnable.
Apache-2.0 · v0.1.0-rc4 · World ranking snapshot ~86% composite (2026-05).