open source · local-first · MIT

Your agent can read `git log`.
It can't read the four things you tried last Tuesday that didn't work.

NexusMem records what actually happened on your machine — shell commands and their exit codes, git history down to the patch of each changed file, project docs, optionally your assistant transcripts — into a local SQLite database, and serves back a ranked, token-budgeted slice of it on demand. Everything stays on disk. No account, no cloud, no telemetry.

Get started View on GitHub

shell

# from inside any git repository
npx nexusmem init
npx nexusmem sync

See it work

Real output, this repo's own history

Init, sync --github, and a query — a real issue, the PR that closed it, and the commits it shipped, recorded against NexusMem's own repository.

Terminal recording of nexusmem init, sync --github, and a query surfacing a real GitHub issue, the PR that closed it, and the commits it shipped, run against this repo's own history

One query, two sources

A commit and a docs section, ranked against each other

Nothing here is summarized by a model on the way out. The ranker just decides what not to send.

nexusmem query

$ nexusmem query "windows spawn failure"

Relevant history for: windows spawn failure

- 2026-08-09 fix: distinguish a failed git spawn from "not a git repository"
  readRepoInfo collapsed three unrelated failures into one error: git
  running and reporting the path is not a work tree, git not being
  installed, and the process failing to spawn at all. Dogfooding hit
  the third case in two separate sessions...

- 2026-08-09 README.md — Before a tagged release
  - [ ] Retry on transient process-spawn failures on Windows

$

What it tracks

Every source normalizes to the same shape

A commit, a shell command and a docs section compete on equal terms — BM25 and vector search, fused by rank.

Shell history + exit codes

A PSReadLine-first hook captures working directory, exit code and a real timestamp. A failed command is a stronger signal than a successful one.

Git commits & diffs

Full commit history plus the patch of each changed file — what shipped, down to the lines that changed.

Project docs

Markdown chunked at heading boundaries. Ask why a decision was made and get the actual rationale section back.

Failure → fix chains

Opt-in. Links a failed command to whatever later resolved it, so its fix rides along in the same result automatically.

Session summaries

Opt-in, local Ollama model, ingest time only. One distilled node — what was decided and why — next to the raw exchanges.

Cross-project recall

query --all-projects searches every repo you've synced, tagged by source. Off by default, scoped when off.

Forget, for real

nexusmem forget <value> deletes every matching node and writes a standing deny-list entry, so it can't be re-ingested by a later sync --rebuild or a fresh clone.

Provenance, staleness & review

Every node carries a trust tier — observed, authored, recorded, derived — that decays at its own rate. A local model checks for contradictions automatically on sync and flags what it finds, dismissible with stale --dismiss if it's wrong. nexusmem review lets a human verify or reject a node directly, independent of the model.

Multi-language import graph

JS, TS, Python, Go, Rust, Java and PHP: parses real import/module statements into edges between files, so "what imports this" is answerable without a repo-wide grep.

GitHub issues & PRs

Opt-in — the first source with a real external dependency. Reads issue/PR threads via the gh CLI, title through every comment, into the same searchable memory.

What it costs you

Two numbers, kept apart on purpose

End-to-end saving compares the packed context NexusMem sends against reading, in full, the files its own ranking identified as relevant.

95%token saving

this repo — 62 commits, 16 real prompts, vs. full file content

99%token saving

vitejs/vite — 9,567 commits, mechanically sampled, vs. full file content

Both clear the original >70% target, measured with a reproducible script (scripts/benchmark.ts) anyone can re-run. The file set each query is graded against comes from NexusMem's own ranking, not an independent judge — this measures what packing saves once retrieval already picked a candidate set, not whether that set was the right one. Full methodology in the README.

For a source-level comparison against similar tools' own numbers, see docs/competitor-comparison.md (vs. projectmem) and docs/competitor-comparison-yesmem.md (vs. YesMem, including native Windows support vs. its documented WSL2 requirement).

Latency, ~530-node corpus, warm p50 over 10 runs
Operation	Time
BM25 retrieval (FTS5)	~1.1 ms
Vector KNN (sqlite-vec)	~3.2 ms
Fuse, rank, pack	~0.6 ms
Query embedding (local Ollama)	~55–77 ms
End-to-end hybrid	~56 ms

Verify it yourself

Pick a real query, see the real numbers

Every result below is replayed output from an actual nexusmem sync + query run against vitejs/vite — the same 16 prompts behind the 99% figure above. Nothing runs live on this page; type or click one to see the recorded numbers, or type list to see every prompt.

PS C:\dev\vite

PS C:\dev\vite>

Replayed, not live-executed — this static site has no backend to run a real query against. The full recorded set is bench/results/vite-final.json, produced by the same scripts/benchmark.ts linked above. To run it for real, against your own repo: npx nexusmem init && npx nexusmem sync && npx nexusmem query "…". Page counts in the output are a rounded estimate (~667 tokens/page, from OpenAI's own ~750-words-per-1,000-tokens rule of thumb) — a way to picture the token counts, not an exact measurement.

Read before you trust it

Where it breaks

Stated plainly, same as the README — including the parts that make the pitch look worse.

Shell history without the hook is unscoped. Scraped history has no directory context and is attributed to whichever repo you ran sync from — a tail-window approximation, not a guarantee.

Japanese and Chinese depend on the vector pass. FTS5's tokenizer splits on whitespace, so languages without space boundaries get no useful BM25 recall.

Rebasing strands nodes. Rewritten history leaves nodes for unreachable commits. A targeted prune doesn't exist yet — sync --rebuild forces a clean re-scan.

Session-summary titles depend on a 3B model following instructions, which it does about a third of the time. The fallback keeps them specific, not elegant.

Cross-project recall favours breadth. Each repo's hits are fused by rank, so a project with only a mediocre match still contributes a result — there's no per-project quality weight.

Try it

Two minutes, inside any git repo

Requires Node 22+ and git. Ollama is optional and only affects semantic search.

CLI

npx nexusmem init
npx nexusmem sync
npx nexusmem query "why did the retry fail"

MCP (agent config)

{
  "mcpServers": {
    "nexusmem": {
      "command": "npx",
      "args": ["-y", "nexusmem", "mcp"]
    }
  }
}

init: Set up .nexusmem/ in the current repository.
sync: Ingest git, diffs, shell history and docs; embed anything new. Add --github to also pull this repo's github.com issue/PR threads for the run.
query <text>: Ranked, token-budgeted context for a question, printed to stdout. Add --as-of <date> to read what the store held at that point, not what's true now.
status: What's currently remembered, by source.
hook install: Wrap PowerShell to capture exit codes and timestamps exactly.
forget <value>: Delete every node matching a value and deny-list it, so it can't come back.
stale: List aging nodes and any contradictions a local model has flagged. Add --check-contradictions to check now, or --dismiss <id> to reject a suggestion that's wrong.
mark-stale <id>: Link a superseded node to its replacement; the ranker down-weights it.
review <id>: Record a human verdict — --verify or --reject — independent of the automatic checker.
mcp: Run the MCP server over stdio for an agent to call directly.