Persistent memory
for AI agents.
Drop in with two lines. Runs in the background.
No database setup. No cloud account.
Every agent you build needs memory from scratch.
No shared infrastructure. No standards. Start over every time.
RAG chunks are large, noisy, and miss relationships.
Vector similarity doesn't capture who recommended what to whom.
Existing graph tools need Neo4j, Docker, cloud accounts.
Heavy infrastructure for a problem that deserves a local file.
How it works
pocket-mem sits silently alongside your agent. Every conversation turn is observed, extracted into a knowledge graph, and stored locally. When your agent needs context, it queries the graph — not a pile of text chunks.
The result is an agent that remembers people, decisions, tools, and relationships across every session.
Watch it work
Feed pocket-mem a message and watch it extract, store, and connect.
answerable questions, Veloris benchmark
across 3 independent domains
contamination detected
Features
Knowledge graph
Not a chunk store. Typed entities and relationships — people, tools, decisions, events.
Zero infra
One SQLite file. Embedding model bundled (~22MB). No server, no cloud, no setup beyond pip install.
Hybrid search
BM25 + vector similarity. Finds facts even without exact keywords.
Visual explorer
pocket-mem show opens a constellation graph in your browser.
Ready to add memory to your agent?
Read the docs →