agentvec

AgentVec

The SQLite for mutable agent memory. A lightweight, serverless, embedded vector database written in pure Rust — built for AI agents that need to remember, forget, and update knowledge in real time.

No server. No cloud. No external dependencies. Just point it at a directory.

pip install agentvec

import agentvec

db = agentvec.AgentVec("./agent_memory.avdb")
memories = db.collection("episodic", dim=384, metric="cosine")

# Store a memory that expires in one hour
memories.add([0.1, 0.2, ...], {"event": "user said hello"}, ttl=3600)

# Recall the most similar memories
for r in memories.search(query_embedding, k=10):
    print(r.id, r.score, r.metadata)

---

Why AgentVec?

Most vector databases are built for search over static datasets — you index once and query forever. Agent memory is the opposite workload: it changes constantly. Agents add new observations every turn, update what they already know, and let stale context decay away.

AgentVec is designed for that mutable, transactional, row-oriented pattern instead of bulk analytics.

Built for

• AI agents — conversational memory, working memory, long-term knowledge
• Local-first apps — no server, no network, just a directory on disk
• Mutable workloads — frequent upserts, deletes, and TTL-based expiry
• Embedded use — single library, low memory footprint, instant startup

Not built for

• Billion-vector or distributed search (use Qdrant, Milvus, or Pinecone)
• Heavy analytical queries over vectors (use LanceDB or DuckDB)
• Static datasets you never update (use FAISS or usearch)
• Multi-process concurrent access (AgentVec is single-process, multi-threaded)

---

Features

• Memory decay (TTL) — give any record a time-to-live; expired memories are skipped on read and reclaimed on compact().
• In-place upserts — upsert(id, ...) is idempotent and ACID-safe; update a memory without rebuilding anything.
• HNSW approximate search — auto-enabled past a threshold, built in parallel across all cores, with incremental inserts (no full rebuilds) and on-disk persistence.
• Metadata filtering — filter by metadata with $eq, $ne, $gt, $gte, $lt, $lte, $in, $nin.
• Multiple distance metrics — cosine, dot, and l2 (Euclidean).
• Crash-safe storage — vectors live in memory-mapped files for zero-copy reads; metadata lives in an ACID key-value store with checksums and a recovery path.
• Product quantization — optional compression for larger collections.
• Export / import — snapshot a collection to a file for backup or transfer.

---

Quick start

import agentvec

db = agentvec.AgentVec("./my_agent.avdb")

# Different memory types can use different embedding dimensions
episodic = db.collection("episodic", dim=384,  metric="cosine")   # conversations
semantic = db.collection("semantic", dim=1536, metric="cosine")   # knowledge
working  = db.collection("working",  dim=384,  metric="cosine")   # scratchpad

# Add a memory (auto-generated id, expires in 24h)
mem_id = episodic.add(vector, {"user": "alice", "type": "chat"}, ttl=86400)

# Update it later — same id, no rebuild
episodic.upsert("conv_turn_42", vector, {"summary": "discussed timeline"})

# Bulk insert is 10-100x faster than one-at-a-time
episodic.add_batch(vectors=[...], metadatas=[...])

# Search, filtered by metadata
results = episodic.search(query_vector, k=10, where_={"user": "alice"})

# Reclaim expired memories and flush to disk
episodic.compact()
db.sync()

---

Companion packages

AgentVec is the storage engine. Higher-level libraries build on top of it:

• agentvec-memory — tiered agent memory (working / session / project / user) with pluggable embedders, so you store text and let it handle vectors and TTLs.
• agentvec-mcp — a Model Context Protocol server that exposes AgentVec memory to MCP-compatible AI tools.

---

Documentation & source

• Repository: https://github.com/Akuming/agentvec
• Issues: https://github.com/Akuming/agentvec/issues
• Full API reference, architecture notes, and benchmarks are in the repository README.

License

Dual-licensed under either of MIT or Apache-2.0, at your option.