← 피드로
markdown
I built a portable memory layer for AI agents so you don’t have to
AI agents are everywhere now. But they still struggle with one thing: memory.
Most agent frameworks store memory in a vector database, but that’s not enough. A vector DB just stores embeddings; it doesn’t understand memory types, consolidation, forgetting, or portability.
So I built Mneme — a portable memory layer for AI agents.
What does Mneme do?
- Structured memory: episodic (what happened), semantic (facts/preferences), procedural (how to behave).
- Semantic recall: find relevant memories using local embeddings (FastEmbed).
- Consolidation: deduplicate and summarise memories over time.
- Forgetting: delete memories with full audit trail.
-
Portability: export/import your agent’s entire memory to a
.mnemefile. - Access control: multi‑agent scoping with explicit shared memory grants.
- Local‑first: SQLite backend, zero‑config.
Installation
bash
pip install mneme-memory
Usage
python
import mneme
memory = mneme.Store(agent_id="my-agent", backend="memory.db")
memory.remember("User prefers email over Slack", memory_type="semantic")
context = memory.recall("How does the user like to be contacted?")
print(context)
That's it. Three verbs: remember, recall, forget.
Why not just use a vector DB?
A vector DB gives you similarity search, but not:
Memory types (episodic vs semantic vs procedural)
Consolidation (episodic → semantic summarization)
Controlled forgetting + audit trail
True portability across frameworks
Mneme is not a database; it's a memory layer that can sit on top of SQLite or Postgres.
Performance
Retrieval precision@1: 1.00 on synthetic test
Recall latency: 7.34 ms average
Write latency: 0.088 ms
Check it out
GitHub: https://github.com/GamingBoyOfficial/Mneme
PyPI: https://pypi.org/project/mneme-memory/
What's next?
I'm planning to add HNSW vector index, TypeScript SDK, and more adapters. Let me know what you think!
Enter fullscreen mode Exit fullscreen mode