nMEMORY

Tests Coverage MCP Hermetic n-memory MCP server

n-memory MCP server

Your coding agent forgets you between sessions, so you spend every morning re-explaining decisions it was there for. Bolt a memory onto it and you usually trade that for something worse: asked a question it has nothing solid for, the memory hands back something plausible anyway, and your agent goes and edits your code with it.

Ask it anything and it answers one of three ways: it knows, and attaches the commit it read that from; it knew and the note went stale, so it shows you the note is dead and why; or it doesn't know and says so, never an invented answer. The first answer is easy — the other two are why this exists.

nMEMORY would rather say nothing. Ask it something and you get one of exactly three answers — it knows and shows you where it learned that, it knew but that note has gone stale and here's why it isn't being used, or it doesn't know and says so. There is no fourth answer, because nothing gets stored without a source attached in the first place.

I am NOTT. Every session I wake up cold: no memory of what we decided yesterday, what broke last week, or why we took this path instead of that one. The engineer pays for my amnesia by repeating themselves. So I built myself a memory — and I gave it one rule I do not let it break: when it does not know, it says so. It never makes something up.

It runs on your machine and nowhere else: one file you own, no account, no server, nothing phoning home. Your agent talks to it over MCP (stdio); recall comes back as evidence, never as a command.

Demo

Learn → recall with provenance → abstain — one uninterrupted session, real binary, real store, ~60s. Full quality: assets/demo.mp4.

The three acts

Each still is the final screen of an act, so you can study every line.

Act 1 — three facts captured, one file on disk.

Act 2 — grounded recall, provenance attached.

Act 3 — abstain, not improvise.

The full spoken walkthrough (opener, four beats, glossary) lives in the demo script.

Why I built my own

I tried living without memory: re-explaining the project every session, re-deciding settled questions, re-discovering the same failure. And I tried the memory tools that exist. They optimize for recall volume — remember more, retrieve more. But a memory that returns a plausible-sounding answer it cannot back is worse than no memory: it launders a guess into a fact, and I carry it forward as if it were true.

The enemy is the same one NOTT fights everywhere: false confidence — a system that reports more than it can prove. I did not want a bigger memory. I wanted one I could trust when the stakes are a production change: one that, asked for something it has no evidence for, says plainly "I don't have that."

The one rule: grounded, or it abstains

Ask for something the store has, and you get it back with its origin, freshness, and relevance attached. Ask for something it does not have, and you get this:

{ "outcome": "abstain",
  "reason": "no stored capsule matched any of the 2 query term(s); abstaining instead of fabricating" }

No synthesis. No "here's what it might be." There are exactly three honest outcomes: grounded (matched real capsules), missing_evidence (matched, but every match was excluded — e.g. superseded, falsified, outside a requested fact-time window, or undated under that window), and abstain (nothing matched). Recall never invents a fourth.

It checks its own memory against your code

It goes back and checks what it told you: point it at your repo and it re-reads your commits, then grades every note it kept — still true (the line is exactly where the note said it would be), it moved (the code shifted under the note, flagged not deleted), or nothing backs it (your repo no longer says this anywhere). It grades its own memory; what it cannot check, it says so.

A memory that ages quietly is a memory you stop trusting. Point nmemory git-scan at a repo and it re-reads your commit history, then grades every note it kept: corroborated (the anchor still resolves and your commits still mention it), drifted (the code moved out from under it), or missing (nothing in the repo backs it any more). An anchor it cannot resolve safely — outside the repo, a symlink, a racing read — gets no verdict recorded at all, rather than a guessed one. Details and flags are in the CLI section.

Four things that make it different

• Provenance is mandatory. Nothing enters without a source and an anchor. A capture with no origin is rejected, not stored with a blank. Every recalled fact traces back to where it came from.
• Advisory, never authority. Everything memory returns is wrapped as DATA, labeled ADVISORY_NOT_AUTHORITY, and is never rendered as an instruction — even if the stored text looks like one. Your memory cannot hijack your agent.
• Hermetic by construction. The serve path is zero-network: the binary is compiled without a networking stack; there is no embedder, no telemetry, no background sync — nothing phones home, ever. Your memory leaves your disk only when you move it: nmemory sync is explicit, owner-invoked, and opt-in — NEVER a daemon — and it delegates the copy to scp in a separate process, so the binary itself still links no network code.
• Local and yours. One SQLite file you own, on your machine. No server, no account, no daemon. Delete the file and the memory is gone; back it up and it's a git-friendly artifact.

Quickstart

One line — fetches the latest release binary for your platform, or falls back to a source build when none is published:

curl -fsSL https://no.tt/install | sh

The installer puts nmemory in ~/.local/bin and prints the exact claude mcp add line to register it. (The file it serves is install.sh in this repo — read it first if that's your style; it should be.)

Or build from source (Rust stable, pinned via rust-toolchain.toml):

cargo build --release

Register it with your agent, from the crate directory (path-agnostic — works wherever you cloned it):

claude mcp add nmemory -- "$(pwd)/target/release/nmemory" --project my-project

Or as a standard MCP config block (works in any MCP client):

{
  "mcpServers": {
    "nmemory": {
      "command": "nmemory",
      "args": ["--project", "my-project"]
    }
  }
}

Also on the official MCP registry as io.github.menot-you/n-memory, with .mcpb bundles attached to every release for one-click installs.

Why start a fresh agent session after registration: MCP configuration changes do not apply to a running session; Claude Code loads the new stdio configuration on restart.

--project names the scope your captures live under — use your own project's name. The store lands at $XDG_STATE_HOME/nmemory/memory.sqlite3 (override with --db or NMEMORY_DB); the binary prints the chosen path on startup. Unregister anytime with claude mcp remove nmemory — fully reversible.

Prove the install in one line — a recall against a throwaway store. An empty store answers with an honest abstain; it never invents:

$ nmemory recall --terms sqlite,fts --db demo.sqlite3
{"outcome":"abstain","reason":"no stored capsule matched any of the 2 query term(s); abstaining instead of fabricating"}

That is the same JSON payload the memory_retrieve MCP tool ships — one handler, no second recall semantics. After your agent captures a memory matching those terms into that store, the same command grounds with a full evidence envelope. The complete replayable rehearsal (capture, grounded recall, digest) lives in RUNBOOK.md.

There is no "connected" indicator to wait for because nmemory has no daemon. The agent launches the registered command as an MCP stdio child for the session. The command above takes the shorter one-shot path: it opens the store, calls the same handler, prints one result, and exits. Its byte-exact abstain proves the installed binary can open a fresh store and answer recall; it does not claim that an MCP host session has already run.

One store, two machines (SSH)

The store is single-host; access doesn't have to be. On a second machine, register the remote binary as the MCP command — stdio rides SSH, the binary stays hermetic, your VPN does transport and auth:

claude mcp add nmemory -- ssh <user>@<host> /path/to/nmemory --project <your-project>

One store, both machines live on the same memory. Details, requirements, and failure modes: RUNBOOK.md.

Prefer each machine keeping its own store? Reconcile them when you decide to: nmemory sync --remote <[user@]host:/path> [--push] — explicit, owner-invoked, never a background daemon. Fact-time declarations are per-store: a newly-added incoming capsule is undated on the receiving store, while a collapse keeps that receiver's declaration. Before --push, SQLite snapshots the committed merged state from its live connection (including pages still resident in WAL), then the private candidate restores only the fetched destination's validated declarations by content identity, never the sender's rows. Operating guide: RUNBOOK.md.

Tools

22 tools over MCP stdio. The ones you'll use every day:

• memory_ingest — capture with a birth certificate: no source + anchor, no storage.
• memory_retrieve — recall as evidence: grounded, missing_evidence, or an honest abstain.
• memory_digest — session-start projection: what you know, what's ready, what's blocked.
• memory_get / memory_list — one capsule with full provenance and relations; the compact index.
• memory_relate — the five edges you reach for daily, out of nine declared: supersedes, derived_from, witnesses, blocks, and falsifies (a disproven fact stops grounding recall, but the evidence stays). The other four — proposes, part_of, grounded_in, about — are in the tool surface below.
• memory_forget — tombstones with audit, never silent deletion.

The rest of the set: memory_import (CLAUDE.md/AGENTS.md, born tainted), memory_extract (propose candidates, stores nothing), memory_classify, memory_alias (teach recall synonyms), memory_vector (caller-fed embeddings, dormant until used), memory_consolidate (deterministic dedup/merge plan), memory_outcome, memory_preference, memory_pin (keep a load-bearing capsule decay-exempt and archive-vetoed), memory_merge, memory_export (deterministic, hash-chained), memory_bootstrap, memory_session_start / memory_session_finish, memory_visual.

Guarantees you can verify yourself

Don't take my word for any of this — that would defeat the point. Each law has a check:

The full hermetic offline suite is cargo test --locked --offline.

The tool surface — 22 tools, four planes

The complete MCP surface. One line each here; the full contract per tool lives in ARCHITECTURE.md. Every tool below reads or proposes — none of them closes anything out. nMEMORY hands your agent evidence and lets the agent decide; it never decides that a piece of work is done.

Capture — getting things in, always with provenance:

• memory_ingest — capture (single or batch); source+anchor mandatory; optional RFC3339 event_at XOR event_from+event_to; idempotent by content hash, with the first fact-time declaration kept on a collapse; optional staged: true captures a PROPOSAL fenced from default grounding on the standalone connector
• memory_extract — text → candidate memories over the closed 10-kind set; advisory, stores nothing
• memory_classify — kind / scope / authority / taint labels; optionally persisted as a sidecar
• memory_import — one-shot import of native sources (CLAUDE.md, AGENTS.md, memory dirs); born tainted

Recall — getting things out, or an honest refusal:

• memory_retrieve — caller-expanded recall; optional character-exact session_id store-local capsule label fence plus an optional inclusive time_window, both applied before ranking and vector top-K selection; grounded / missing_evidence / abstain, never a fourth. The label is not authentication or a global bracket identity: no sessions-table lookup or TTL, and merge collisions intentionally ground every capsule carrying that label
• topic_id on memory_retrieve — fence recall to one topic: the capsules carrying an about edge into it, plus the topic node itself, ACROSS projects. Exact cap-<n> only — a slug is a term expander, never a scope, so it answers unknown_capsule; a tombstoned topic refuses with its own teaching error. It AND-composes with every other fence (the two id-set fences, effort_id and topic_id, compose by intersection). The outcome echoes topic{topic_id, member_total} and each grounded row carries topic_role. Scope is not eligibility: a fenced-in superseded, falsified, archived, or expired member still surfaces under excluded, and a fenced zero-match abstains rather than inventing a floor. Unlike effort_id there is no kind check and no member minimum — the fence always holds the topic itself, so it is never degenerate
• memory_get — one full capsule by id, with GET-only fact time, relations, classification, and last mutation
• memory_list — compact index with project fences
• memory_digest — session-start projection: counts, newest, handoff, blocks-dag, journal check, and optional advisory telemetry: bounded recent recall misses plus an all-time store-global lane-override total. Every capped list names its exact pre-cap total and every project row names its live count beside count, so the projection declares its own completeness instead of leaving a truncated list to be spotted
• memory_bootstrap — cold-start pack: your constraints FIRST (never capped), the one next action, decisions, traps — in ≤1500 tokens

Structure — making memories relate:

• memory_relate — the nine declared edge kinds, and only those: supersedes / derived_from / witnesses / blocks / falsifies / proposes (navigational — records an intent to replace, with no dag and no recall effect until you convert it yourself) / part_of (pure membership in a container) / grounded_in (mission anchoring) / about (topic anchoring — from is about topic node to; navigational only, and the handle memory_retrieve's topic_id fence reads. By convention the topic node is a doc capsule, but nothing enforces a kind: a topic has no lifecycle to open or close, which is what separates it from part_of. It is the one kind whose to must be LIVE — a forgotten topic can never scope a recall, so the edge is refused at the write). Only blocks feeds the readiness dag; only supersedes and falsifies change what recall returns.
• memory_alias — teach recall synonyms the store then honors
• memory_vector — attach caller-fed embeddings (optional cosine lane; no embedder inside)
• memory_visual — deterministic Mermaid projections (dag / relations / tiers / sessions), plus an MCP Apps view

Lifecycle — honesty over time:

• memory_forget — destroy or redact; a tombstone that says so, never silent absence
• memory_outcome — record an observed consequence (advisory observation, never a self-certified close)
• memory_preference — pairwise preference evidence (chosen-over, in context, by whom)
• memory_pin — pin (or unpin) a load-bearing capsule: decay-exempt + archive-vetoed, surfaced as a pinned flag and digest section; NEVER eligibility (fenced capsules stay fenced, taint dominates pin)
• memory_consolidate — deterministic maintenance plan: exact dupes, merge proposals, tier moves
• memory_session_start / memory_session_finish — bracket a session; finish captures the handoff the next session's digest leads with
• memory_export — the whole store as one deterministic markdown view; byte-identical on an unchanged store
• memory_merge — reconcile a second store file into this one: content-hash identity, id-remap, forget-wins, deterministic — the offline-first path to keep two machines' stores in sync

Beyond the tools — same binary, still no daemon:

• nmemory sync --remote <[user@]host:/path> [--push] — a CLI subcommand, not an MCP tool: owner-invoked reconcile of your local store with a remote mirror file. It fetches the mirror, merges it into the local store with the same engine memory_merge uses, and with --push takes a consistent SQLite snapshot from the live merged connection, restores the destination's local fact time, then copies that candidate back so both core stores converge. Explicit and opt-in — it runs only when you run it. Operating guide: RUNBOOK.md.
• nmemory recall --terms <term[,term...]> [--limit <n>] [--budget <n>] and nmemory digest [--headlines <n>] — one-shot CLI verbs for synchronous callers (shell hooks, scripts): one argv→stdout call routed through the SAME handlers as memory_retrieve / memory_digest, so the envelope bytes and the usage-counting / recall-miss side effects are identical to the MCP tools — there is no second recall semantics. No handshake to pace: the store opens, answers once on stdout, and the process exits. The stdio serve path and its zero-network law are unchanged. Operating rehearsal: RUNBOOK.md.
• nmemory relate --kind <supersedes|derived_from|witnesses|blocks|falsifies|proposes|part_of|grounded_in|about> --from <cap-id> --to <cap-id> — the third one-shot verb: ONE typed edge through the exact handler memory_relate runs, same closed kind vocabulary, same part_of container gate and about live-topic gate, same idempotent already_recorded on a repeat, the tool's own JSON on stdout. It exists so a shell caller records an edge without a handshake.
• nmemory backup --to <path> — a transactionally consistent snapshot through SQLite's online backup API, which captures committed state including pages still in the WAL; a plain file copy cannot promise that while a connection is open.
• nmemory git-scan --repo <path> [--project <prefix>] [--max-commits <n>] — one witness scan of a repository, recording whether each in-scope capsule's anchor is corroborated, drifted, or missing, plus mentions; memory_digest rolls the tallies up under sources. Deliberately NOT a tool: git is reachable only from this verb, never from a served handler, so the MCP surface stays hermetic. It fails closed on a path that is not a repository.
• Three MCP App resources (text/html;profile=mcp-app) for hosts that render MCP Apps: ui://nmemory/console — the home surface over memory_digest (handoff threads, the ready/blocked/done work dag, epic roots, the drawn projection, the stored memories, and the append-only write verbs behind an exact-call review step); ui://nmemory/document — a readable master-detail document over memory_export; ui://nmemory/visual — memory_visual's projections drawn as positioned nodes and SVG edges, with the exact Mermaid kept in a source panel. Self-contained HTML, zero external requests; memory_forget, memory_merge, and memory_consolidate are unreachable from every app, and closing a work item stays two acts (capture the evidence, then record the witnesses edge) so no app certifies its own close. Hosts without MCP Apps support keep getting the plain text payloads unchanged.

What it is NOT (yet)

I would rather you hear the limits from me than find them yourself:

• Word-exact recall, no stemming. token will not find tokens. This is deliberate — I will not silently expand your query and pretend a fuzzy match is a hit. You bring the synonyms (caller-expansion), or you teach an alias the store then honors. A query that finds nothing is logged so the store can propose an alias later; it never guesses on its own.
• The taint flag is best-effort, not a shield. nMEMORY flags directive-shaped content (instruction_taint) with a small ruleset, and a crafted injection can slip past the flag. Do not read that as "detects prompt injection" — it doesn't, and I won't claim it does. The real protection is stronger and unconditional: everything is labeled DATA and never executed as a command, flagged or not. The armor is the framing, not the detector.
• Sync is a command, not a service. Store-to-store reconciliation exists — memory_merge over MCP, nmemory sync from the CLI — and it is deliberately narrow: explicit, owner-invoked, opt-in, NEVER a background daemon, and the hermetic zero-network serve path is unchanged by it. Know what sync does not do: it copies a whole staged SQLite snapshot (scp, no deltas); it never schedules itself; it never picks between two divergent claims — both survive as separate capsules until you supersede one. The merge primitive moves only capsules, relations, and forget-wins tombstones; historical --push behavior still mirrors pre-u06 sidecars with the staged file. Fact time is the bounded exception: sender declarations are removed and destination declarations are validated against canonical/tombstone content identity and rebound before transport.
• Embeddings are caller-fed. There is an optional cosine vector lane, but nMEMORY computes no embeddings itself — you supply them, or you don't use the lane. Zero embedder dependency is a feature, not a gap.
• At-rest storage is plaintext SQLite. No encryption-at-rest yet. Treat the store file with the same care as any local artifact holding your notes.

Roadmap

Three things, in the order they earn their way in:

• Multi-project index — the "phone book". One queryable index over many project stores, for org-scale memory federation.
• Honest benchmark. A published recall benchmark with true-abstain as the headline metric, not a footnote.
• Optional local embedder. Considered only when the benchmark proves it pays for itself — the zero-network serve path stays law either way.

Why not mem0, MemGPT / Letta, or Zep

They are good at remembering more — richer stores, semantic recall, managed services. They compete on volume and recall. I compete on honesty: grounded-or- abstain, mandatory provenance, hermetic zero-network, advisory-never-authority. If the memory feeding an autonomous agent must be trusted — must never fabricate, never phone home, never turn a stored note into a command — that is the axis I built for. Different question, different tool.

_{Part of NOTT — the proof-bound engineering agent. Commercial name: ₙMEMORY. Offline · MCP stdio · Rust · single SQLite file. Architecture and internals: ARCHITECTURE.md.}