feat: add Mimir — persistent agent memory engine (FTS5 hybrid, no embeddings)

[tcconnally]

Mimir: Persistent Agent Memory Engine

Mimir is a persistent agent memory engine with structured entities and FTS5 hybrid search. It achieves 72% accuracy on LongMemEval oracle split — beating FullContext by 26 percentage points — while using no embedding API.

Key Features

• ✨ Structured entity storage with tags, categories, decay, links, verification
• 🔍 FTS5 hybrid search — semantic keyword matching without embeddings
• ⚡ HTTP/SSE MCP transport — connects to a running mimir serve instance
• 💰 Zero embedding cost — all retrieval is local SQLite FTS5
• 🔗 Proactive recall hooks (recall_when) for agent-driven memory

LongMemEval Results (oracle split, 50 temporal reasoning questions)

System	Accuracy	F1	Tokens
Mimir	0.720	0.201	277,824
FullContext	0.460	0.285	381,439

LLM: gemma-4-26b-a4b-it via OpenRouter

Compared to Published Systems (different LLM/judge)

• Mimir (0.720) matches PropMem (0.716) on accuracy while using 82× fewer tokens
• Mimir (278K tokens) is 2.5× leaner than OpenClaw (0.7M) on comparable F1

Setup

# Start Mimir
mimir serve --db /tmp/mimir_bench.db --transport sse --port 8766 &

# Run benchmark
uv run python scripts/run_full_benchmark.py \
  --benchmark longmemeval --split oracle \
  --systems mimir --num-samples 50

Links

• Repo: https://github.com/Perseus-Computing-LLC/mimir
• Docs: https://perseus.observer/mimir
• Install: curl -sSf https://perseus.observer/mimir/install.sh | sh