agent-memory-weaver

A small research scaffold for multimodal long-term memory in LLM agents.

Motivation

As agent systems move from single-turn interaction to persistent workflows, memory quality becomes critical. This project explores a practical memory stack with:

• deduplication
• recency-aware retrieval
• lightweight salience signals

Features

• unified memory item schema (text + optional image_caption)
• near-duplicate filtering
• recency decay + semantic similarity scoring
• timeline-friendly retrieval traces

Install

python -m venv .venv
source .venv/bin/activate
pip install -e .

Quickstart

python -m agent_memory_weaver.cli add \
  --store examples/store.json \
  --id m1 \
  --text "Compared two VLM checkpoints on OCR-heavy samples" \
  --tags eval,ocr

python -m agent_memory_weaver.cli search \
  --store examples/store.json \
  --query "OCR benchmark insights" \
  --top-k 3

Scoring

Retrieval score is a weighted sum of:

• semantic similarity (hashing vector cosine)
• salience (user provided, default 0.5)
• recency decay (exponential, configurable half-life)

Roadmap

• Add cross-session memory graph links
• Add modality-specific salience heads
• Add memory compaction and snapshotting

Notes

Best used as a prototype before integrating production vector DBs.