mhn-ai-agent-memory

Associative Memory for AI Agents Using Modern Hopfield Networks

A Python library for deterministic associative memory in AI agents.

Status: toy / research memory. Deterministic associative recall for small agent fact stores and learning Modern Hopfield networks. Useful for Cursor project working memory and offline experiments. Not a production replacement for vector DB + LLM memory systems.

Experiments: EXPERIMENTS.md — 43 tests, retrieve latency (~16 µs/op at dim=256), reproducible benchmarks/ (see repulsive A/B v2).

What is this?

Many agent memory systems store text in a database, then ask an LLM to interpret it at query time. This library does something else: it applies the Modern Hopfield Network update rule directly, exposing transformer-style attention as an explicit, controllable memory.

• Store text in a database
• Call an LLM to reason about it (costs money, takes seconds, non-deterministic)
• Hope the LLM doesn't hallucinate about what it stored

Store a fact as a pattern in the energy landscape. Query with a partial cue, and the network relaxes toward the nearest stored pattern. The retrieval weight gives you a confidence signal.

30-Second Example

No API keys, database, or config files. The memory is a NumPy array.

Features at a Glance

Pluggable Encoders

Choose an encoder based on quality, cost, and dependencies.

Contradiction Detection

Flag likely conflicts before storing new facts.