← All selected work

Research Memory Lab

A research workspace that keeps cited evidence, reviewed findings and their history

Local research tool · 2026 Source
github.comOpen full size ↗Research Memory Lab experiments page comparing no context, source retrieval, and sources plus memory for one research question

The problem

Research assistants answer fluently but forget what was checked last week, and rarely show which passage supports which claim. This project keeps evidence, human verdicts and findings over time, and measures whether that memory changes the answers.

My contribution

Personal project: ingestion with page references, hybrid retrieval, Graphiti graph memory, the study-material studio, human review and promotion of findings, and the experiment bench.

  • FastAPI
  • SQLite FTS5
  • Ollama embeddings
  • Graphiti + FalkorDB
  • React 19

What I built

The approach

Import papers and notes, ask questions over them with citations, turn sources into study material, record a human verdict on each claim, and keep versioned findings in a graph memory. An experiment bench compares answers with no context, source retrieval, and retrieval plus memory.

How the comparison is set up

0context conditions compared per question
6 × 600snippet budget, same for both retrieval conditions
Local or hostedgeneration, same evidence

The experiment bench measures whether researcher-selected sources were retrieved, not whether answers are correct. A held-out question set with citation-support labels and repeated trials is still needed before claiming memory improves answers.

Follow the workflow

PDFs, notes and dated web snapshots are imported explicitly, keeping page references.

More from the project

Select an image to inspect it at full size.

Engineering decisions

01

Combine keyword and meaning search

SQLite FTS5 and local embeddings are ranked separately and merged with reciprocal rank fusion.

02

A person decides what is remembered

Only a claim whose latest human verdict is "supported" can become a finding, and revised findings keep their earlier versions.

03

Make the comparison fair

Both retrieval conditions get the same snippet budget, and exports keep context, prompt and model hashes so a run can be inspected later.

Limitations & next steps

Build a labelled question set, run repeated and counterbalanced trials, and report whether memory improves citation support rather than only retrieval.