Travels with the code
Memory is committed alongside the work it describes. Clone the repo and you get the decisions, the conventions, and the gotchas — no separate database to provision, no export to chase.
Repo-resident memory
Your memory is plain files in your own git repo.
No proprietary database, no cloud to trust, nothing to export. Every note, every relation, every provenance stamp is markdown under .memory/ — versioned with the code it describes and read by every agent that opens the repo.
Reviewed like code
Every note is plain markdown with provenance. It diffs in a pull request, gets commented on, and merges through the same review your team already trusts.
Yours if we vanish
Nothing is locked inside the app. The files work in any editor, the CLI, or another agent — and the Heirloom License opens the source if Memophant is ever abandoned.
The shape of a memory
Observations and relations, provenanced to a commit.
Each note pairs atomic observations with explicit relations, so facts are searchable and links are something you authored. A note can record the source files and the commit it was reviewed against — which is what makes drift detection possible later.
- Plain markdown with YAML frontmatter — grep it, diff it, edit it anywhere
- Observations carry tags; relations link notes by name
- source_paths + source_sha let a note know when its code moves
- The app is a fast native UI on top — git stays the source of truth
Built for speed
A search engine that runs inside the app — and leaves the old tool behind.
Memophant used to shell out to basic-memory, a Python tool. On 2026-06-06 a pure-Swift, in-process engine replaced it as the sole backend — hybrid full-text (BM25 over SQLite FTS5) fused with on-device MiniLM embeddings. It runs locally, with nothing to install.
13–113×
faster than the basic-memory tool it replaced
100%
of benchmark queries returned a relevant hit
250
queries across 7 real projects, in one run
0
external processes — a pure in-process Swift engine
Across 250 queries on 7 real projects, the engine returned a result on every one, at a 13–113× speedup. On one head-to-head query it surfaced 20 relevant matches where the old tool found 5 — with the top five identical. Numbers are from Memophant's own benchmarks; your mileage varies with corpus size.
Hybrid search
Lexical precision and semantic recall, fused.
BM25 catches the exact term; the embedding model catches the idea you phrased differently. Memophant fuses both rankings so a typo, a synonym, or a natural-language question all land on the right note — in milliseconds, on your machine.
- BM25 over FTS5 for exact-term and identifier matches
- On-device MiniLM cosine similarity for meaning
- Exposed to every agent over MCP — searched, not stuffed into context
- No Python, no daemon, no network call to search
Memory that outlives the session — and the developer.
Your repo becomes the source of truth, and every agent reads from it.