// tool · free

Keep editable core memory and an append-only archive.

A free Node and Python memory layer with shared files, keyword-and-recency recall, and an optional reranker you supply.

flux@store
$ install memory
✓ resolving dependencies
✓ ready — 0 setup, runs on your machine
› launching…
flux@store: ~/run
$ memory --run
✓ write core or archive entries
✓ prefilter by keyword and recency
✓ optionally rerank candidates
› return the selected memory
done — output ready ▍
$ run --pipeline
1
Write core or archive entries
2
Prefilter by keyword and recency
3
Optionally rerank candidates
4
Return the selected memory
$ explain --how-it-works

How it actually works

Exactly what it does — and when.

1Maintain an editable core plus append-only archive
2Prefilter recall by keyword and recency
3Optionally rerank through an adapter you supply
In practice · A restarted local agent recalls project context

The Node process restarts, reads the shared local memory files, and retrieves recent keyword matches without a model call. The operator inspects the results and later adds an authorized reranker for a larger archive, while keeping access control and retention outside the library.

$ cat spec.txt
IncludedNode and stdlib Python twins, shared file format
SetupChoose a local memory directory
DependenciesZero; reranker optional
BundleFree standalone and in the $199 bundle
$ benchmark --vs-alternatives
This product
A hosted memory service
Storage
Local editable files
Remote managed store
Fallback
Keyword and recency only
Often requires embedding access
Reranking
Optional adapter you supply
Vendor-specific
Boundary
Retrieval depends on your data and query
No guarantee of complete recall
$ diff --before-after
Without it
  • Every restart begins without context
  • Recall requires a paid vector service
  • Current facts overwrite historical evidence
With it
  • Keep a small editable core beside an append-only archive
  • Recall by keyword and recency without an API key
  • Add an optional reranker without changing the stored format

Use it to…

Real ways people run MEMORY — Self-Editing Agent Memory.

Store durable project context

Write current core facts while retaining older entries in the append-only archive.

Recall without a model

Use keyword and recency ranking when no reranker is configured.

Add a custom reranker

Plug in your authorized model or scorer after the local prefilter.

What's inside

Self-editing core memory + append-only archive
recall() = keyword/recency prefilter → optional bring-your-own reranker
Degrades to keyword-only with no API key
Node + stdlib-only Python twin, same data files
Free · zero-dep · MIT source on GitHub

Who it's for

Agent developers

need simple inspectable memory files

Local-first builders

want recall that works without a provider

Node and Python teams

need both runtimes to share the same data

Questions, answered

✓ Free · no email✓ Yours to own✓ Use on every project✓ Built to ship results

Ready to put MEMORY — Self-Editing Agent Memory to work?

Free download · no account needed · yours forever

MEMORY — Self-Editing Agent Memory