Agent memory source, downloads and resources
A small, runnable Python reference for practical agent memory — not a production system, not a research benchmark.
Scope — This is a compact reference implementation (~500 lines) to illustrate core memory patterns: provenance, retrieval, consolidation, and evaluation. It is intentionally minimal and not a complete production memory service. The evaluation fixture is a synthetic corpus of 16 entries and 5 queries, used to verify the pipeline runs — not a research benchmark.
Quickstart
Clone the repository, then run the examples from the /code/examples/ directory:
python quickstart.py
python provenance.py
python -m agentmemory.eval --seed 7 --k 5The evaluation writes /evaluation/eval-results.json with per-query and aggregate metrics. In local tests (Python 3.14, 2026-10-25) the pipeline exits 0 and produces stable aggregate rows; full JSON output is not bit‑reproducible across runs due to unverified ordering differences.
Source Modules
All code is MIT‑licensed and browsable under /code/agentmemory/:
Download & License
The complete reference source is available as a zip archive (MIT license). The archive is generated by the build process; the path is /downloads/agentmemory-reference.zip. The license text is at /code/LICENSE.
Evaluation Artifact
Inspect the latest run results at /evaluation/eval-results.json. The fixture is deliberately small — 16 synthetic entries, 5 queries — to keep the reference fast and easy to understand. Metrics include mean recall@k, mean precision@k, and MRR.
Python Standard Library References
The implementation relies on standard library modules. For deeper context, see the official documentation:
Image Credit
Hero photograph: “Nikon D4 HTTP Server Working… LOL” by Takashi Hososhima, CC BY‑SA 2.0. Source: Wikimedia Commons, license CC BY‑SA 2.0. The image has been cropped and re‑encoded for this site; the derivative is also CC BY‑SA 2.0. It is illustrative of networked hardware, not a depiction of our lab.
.jpg){kind=link}
.jpg){kind=link}