przm Memory
Technical handle: engram·npm: @onenomad/przm-memory
A memory system for AI agents that actually works. Runs locally, doesn't need an API key, and scores 96.8% R@5 / 98.8% R@10 on LongMemEval and 85.1% R@5 / 92.0% R@10 on LoCoMo with a deterministic 9-stage hybrid search pipeline. No LLM in the retrieval loop. Methodology is reproducible from a fresh clone with one command.
// benchmark numbers
Reproducible from a fresh clone with pnpm bench:locomo. Committed result JSON lives in benchmarks/results/published.
// what's in the box
9-stage hybrid search
Vector similarity, BM25 keyword, temporal inference, knowledge graph, spreading activation. Each stage catches what the others miss. Deterministic. No LLM in the retrieval loop.
Knowledge graph
Auto-extracts triples (subject, predicate, object) from ingested content. Graph-aware reranking surfaces related memories that pure embeddings miss. Persists across sessions.
Procedural rules
Learned 'always do X' or 'never do Y' rules from corrections. Survive sessions. Sync with przm Voice via the procedural bridge when both servers are running.
Handoff protocol
Pre-compact and stop hooks write structured 'where we left off' snapshots so context compaction never costs you the session. Resume by name later.
Cognitive layers
Episodic, semantic, procedural: same model as human memory. Each layer has its own decay rate so a one-off fact and a learned habit age differently.
Self-organizing
Duplicate detection, near-duplicate merging, importance decay, cross-link generation. The memory store maintains itself; you don't have to curate.
// install
claude mcp add engram -- npx @onenomad/przm-memory{
"mcpServers": {
"engram": {
"command": "npx",
"args": ["@onenomad/przm-memory"]
}
}
}git clone https://github.com/OneNomad-LLC/przm-memory.git
cd przm-memory
npm install
npm run buildFull installation docs including hooks setup, slash commands, and cloud mode at the GitHub README.
// pairs well with
If przm Memory is the brain, przm Voice (persona) is the personality. Memory handles what the agent remembers. Voice handles how it communicates. They sync via a procedural bridge so frustration spikes get encoded as high-importance memories and applied evolution proposals turn into learned rules.