Memory Path Engine: replayable evidence paths for agent memory retrieval

[ly85206559]

I just open-sourced Memory Path Engine, a research-first prototype for structured memory retrieval in AI agents.

Most retrieval pipelines still stop at top-k chunks: split content, embed it, fetch the nearest matches, and leave the actual reasoning implicit. This project explores a different shape. Instead of treating memory as a flat index, it models memory as typed nodes, edges, and weights, then returns a MemoryPath: a stitched answer plus an ordered, replayable evidence path with per-step scores and edge types.

If you want a more intuitive mental model, you can think of it as a memory-palace-style idea for agents: memory should feel navigable, not like a pile of unrelated chunks.

The goal is not to hide more logic inside the model. The goal is to make retrieval easier to inspect, compare, and debug, especially for multi-hop questions.

What is in the repo today:

• multiple retrieval modes in one codebase: lexical baseline, embedding baseline, structure-aware traversal, weighted graph retrieval, and activation spreading experiments
• replayable path output instead of only final answers
• bundled runbook and contract demos
• repository-owned benchmark fixtures for structured retrieval experiments

Quick start:

python -m pip install --no-build-isolation -e .
python -m memory_engine.demo --scenario runbook
python -m memory_engine.demo --scenario contract

The contract scenario prints a baseline block and the path-aware section side by side so you can compare evidence shape.

If you are interested in agent memory, graph-aware retrieval, or evidence-backed multi-hop reasoning, I would love feedback.

Repository (clickable in GitHub): https://github.com/ly85206559/memory-path-engine

MIT License.