mushroomdb

The graph that stays true — and knows who's allowed to see it.

mushroomdb is the data layer for agents that reason over entities. It is an embedded Rust graph database in which a relationship is a schema declaration: write a rule once, and every write derives, maintains and retracts the matching edges, each one carrying the rule, the score and the values that produced it. An agent reaches it over MCP — twenty-three tools on an entity store — or you embed it as a Rust library, a Python module, or a sidecar beside your own service. Four questions are what it exists for: why are these two related (explain_association, answered with the evidence rather than an assertion), what did that look like then (edges_at, the edges a node had on any past date or commit), what would this change do (what_if, computed without writing anything), and who may see it (query with a role, so one graph answers differently per caller). Local-first: a directory on disk, no account, no endpoint, no model call in the write path unless you enable embeddings.

ingest-git stays supported as a data source: commits, pull requests, files and authors become entities with rule-derived relationships, which is what makes a ticket↔commit link a rule rather than a script.

Removed in 0.7: the code-graph door — seven tools, four hooks and ten subcommands. To keep it, pin mushroomdb@0.6.x; the changelog lists what went.

Pre-1.0 alpha — APIs and formats may change between minor versions.

Docs · Changelog · Issues

A single SET changes one property; the rule fires, new scored edges are derived and the stale ones retracted, inside the same write

Quick start

npx mushroomdb install --db ./memory

One command writes the /mushroom skill, an MCP server listing the twenty-three-tool association surface, and the session hooks. Then the loop a session runs:

remember  →  recall  →  suggest_rules  →  explain_association
 (a fact)    (find it)   (the store        (why two things
                          proposes a rule)   are related)

remember takes a sentence and the entities it names; a subject it has never seen is stubbed rather than refused. recall answers a question in ordinary words. Once a dozen entities carry a field whose values repeat, suggest_rules proposes the rule that links them — nothing is created until you approve its create_rule — and from then on every write maintains those edges. explain_association says which rule linked two entities, and on what evidence.

Full tool reference: docs/site/mcp.md.

• Live, not a snapshot. One SET on a property re-derives the matching edges — added, scored and retracted — before the write closes. The SET in the GIF above is that one write.
• Retracts instead of going stale. A property drifting out of a rule's predicate doesn't leave a stale edge behind; the engine retracts it in the same write that caused the drift.
• Explains any link. explain_association names the rule, the score, and the values the two entities actually share, so "why are these two related?" has an answer your assistant can quote instead of a guess.
• Knows who's allowed to see it. Pass a role or a mask with a query and the same graph answers differently per caller; write statements are rejected on masked queries.
• Answers what it said last week — by date. edges_at("talent-1", "2026-06-19") returns the edges a node had on that day; a 0-based commit index works too, and mushroomdb asof ./db --at 2026-06-19 --query "…" replays the WAL to the last commit at or before it, derived edges included. A store that records no times says so by name rather than guessing a commit.

Agent memory

Graph structure captures the shape of real knowledge — entities, associations, similarity, and lineage — and rule-derived edges keep those associations fresh as new facts arrive.

• Entities map to nodes (Person, Document, Project, Concept, …).
• Associations are edges derived from data: cosine similarity on embeddings, shared field values, FK relationships, geographic proximity. Declare a rule once; every write maintains the matching edges without agent-side bookkeeping.
• Recall has four modes: recall for a question in ordinary words, ranked over every indexed text field; find_similar by query vector (HNSW when available, brute force otherwise); find_similar by key (neighbors along a rule-derived edge type); and query for structured Cypher. hybrid_search fuses fulltext and vector results via Reciprocal Rank Fusion, and pairwise_similar answers "which of these are most like each other" over a caller's own key set, exactly — it never uses HNSW. find_similar's min defaults to 0.8 on every surface.
• Explanations are built in: explain_association shows which rules and scores produced each link, so an agent can cite evidence instead of asserting a conclusion.
• Scoping is a property of the handle, not of each call: db.scoped(role=…, namespace=…, keys=[…]) returns a read-only child sharing the same store, and every read on it obeys one contract — hidden behaves exactly as absent, so a key outside the scope is indistinguishable from a key that does not exist. Legs intersect, so a scope narrows and never widens; every mutation raises ReadOnly. Per-call mask: [key1, key2, …] on query still works and still rejects writes. Both are cooperative in-process argument handling rather than an access boundary — real enforcement is the HTTP server's role tokens. docs/site/masks.md
• Bulk loading is one atomic frame: ingest_batch(nodes, edges, on_conflict="error" | "skip" | "replace") lets a mirror rebuild onto a store that already has content without wiping the directory, and reports inserted, edges_inserted, skipped, replaced and kept_view_owned. "error" is the default and is the older behaviour exactly.
• Schema-as-code: mushroomdb schema apply <dir> <schema.json> idempotently applies rules, views, and fulltext indexes, printing a created/updated/unchanged diff. Two presets need no file: --memory-defaults gives an existing store the text fields recall searches, and --memory-identity adds the SAME_AS rules that link two keys for one entity, after saying what it will backfill.

Eleven task tools answer a question in prose in one call, on any store. They are what the skill reaches for. mushroomdb mcp <db> --all-tools lists these eleven first; the default listing mixes them with the graph tools and opens with query and explain_association:

Each of the eleven also takes json: true, which answers with the raw report instead of the rendered digest.

The fourteen graph tools reach the store directly. Their descriptions are prefixed Advanced: in tools/list, so an assistant knows which surface is the front door. Every store lists the same twenty-three, a store built by ingest-git included — the association surface: query (with an optional role), explain_association, neighborhood, node_info, node_edges, was_linked, edges_at, what_if, node_history, edge_history, find_similar, pairwise_similar, hybrid_search, remember, recall, upsert_entity, ingest_json, create_rule, stats, schema, analyze, suggest_rules and forget. All 25 stay served either way — the listing decides what a session can call, not what the server answers — and mushroomdb mcp <db> --all-tools lists the whole set:

Full walkthrough, tool reference, and Claude Desktop setup: docs/site/mcp.md. Skill, plugin, and hook details: docs/site/skill.md.

Install

Pick the row for what you want to do. The paths are not interchangeable: the last column is what each one leaves out.

Docker details, install.sh, and building the binary with the UI embedded are in CONTRIBUTING.md.

install writes an MCP entry that runs npx -y mushroomdb@<version>, so the assistant needs nothing installed globally and nothing is copied into your home directory. Point it at a local build with --command <path>. mushroomdb doctor verifies the result end to end — config entry, store, lock, hooks, and a real stdio handshake with the configured command.

To see the bundled explorer, write a demo graph and serve it. These lines assume a mushroomdb binary on your PATH that embeds the UI — a release binary, or the one install.sh fetches. The plugin and npx rows put nothing on PATH: there, prefix each line with npx.

mushroomdb demo ./db
mushroomdb serve ./db

Open http://127.0.0.1:8080/. The demo graph has 10 Orgs, 20 Projects, 30 People, and 334 edges — 304 of them derived by seven rule sets. When a token is configured, open http://host:8080/?token=….

Role-bound tokens limit a caller to a named subset of nodes. Define roles in schema.json under the roles key (each role has a label selector list), then pass --role-token TOKEN:ROLE (repeatable) when starting the server, or set MUSHROOMDB_ROLE_TOKENS="tok1:role1,tok2:role2". A role token receives only the nodes matching its label selectors — read endpoints return rows filtered to the visible set; write, subscription, and analytics endpoints return 403. Unknown token or role name: 401. The never-widen invariant is enforced in the server: a client-supplied mask is always intersected with the role mask. The MCP interface (mushroomdb mcp) is a stdio JSON-RPC server for local agent use and is not subject to bearer-token or role enforcement.

Where it fits

What it is

• An embedded, single-binary graph database with a rule engine that maintains edges for you.
• A 25-tool MCP server — twenty-three listed on every store, a store built by ingest-git included — plus a /mushroom skill and a Claude Code plugin.
• Safe for several processes at once: one writer at a time behind an advisory LOCK file, any number of readers, and every handle picks up a peer's commits by refresh() rather than reopening — so a running serve, the session hooks, an ingest-git run and other CLI commands can share one store. docs/site/concurrency.md
• Local-first: your data stays on disk, no cloud service, no model call in the write path unless you enable embeddings.

What it isn't

• Not a hosted memory service — there is no account, no endpoint, nothing to sign up for.
• Not a vector database. Vector predicates and HNSW are built in; bring your own embeddings.
• Not a transactional relational database. Single writer, no interactive transactions, memory-first storage.

The differentiator

Most graph databases require you to create edges manually or run a batch similarity script after each load. mushroomdb makes edge creation a schema declaration. A rule like "connect every Person to every Org whose skills list overlaps theirs by at least 50%" is written once:

db.create_rule(RuleDef {
    name: "skill_fit".into(),
    src_label: "Person".into(),
    dst_label: "Org".into(),
    predicate: Predicate::Overlap { field: "skills".into(), min: 0.5 },
    edge_type: "FIT".into(),
    weight_prop: Some("score".into()),
    max_edges: Some(5), // keep the 5 best-matching Orgs per Person (top-k per source)
}).expect("rule");

After that, every insert_node and set_prop evaluates the rule incrementally. The engine writes the edge, stores the Jaccard score, and retracts the edge if the properties later diverge — without any manual work.

Watch it live — a Cypher SET changes one property, the founded_within rule fires, and new scored edges appear in the bundled explorer:

Open the Rules panel, and the Why slide-over shows the exact predicate arithmetic behind every derived edge:

Rules panel and Why slide-over showing overlap arithmetic

Predicates

Six predicate kinds ship today. They compose via All(...) (AND, score = min) and Any(...) (OR, score = max), nested up to depth 4.

Auto-FK: fields ending in _id whose values match existing node keys get KeyMatch rules created automatically at ingest time. VectorSimilar accepts approximate: true to switch candidate selection to in-tree HNSW (per-query recall floors min 0.90 / mean 0.95, measured 1.0 / 1.0 at 5k nodes / dim 1536, fixed-seed probe). Full reference: docs/site/rules.md.

Built on the same engine

• Live subscriptions. subscribe_rule (Rust) and GET /subscribe (WebSocket) stream EdgeFired / EdgeRetracted the moment they hit the WAL — not polled, not batched. Bounded 65,536-event queue; slow consumers get a Lagged { missed: N } marker instead of a disconnect. docs/site/subscriptions.md
• Rule attribution across time. Every derived edge writes a HISTORY-MARKER WAL record carrying the rule name, so edge_history, node_history, and was_linked answer which rule created a link and at which commit. GraphDb::open_at(&dir, 5) replays to a past commit, derived edges included; out-of-range commits return CommitOutOfRange, never wrong data. docs/site/timetravel.md
• Materialized views. Degree counts and neighbor aggregates (sum/avg/min/max) maintained incrementally on every edge change — no cron, no triggers, no stale caches. docs/site/views.md
• Rule suggestions. db.suggest_rules() (or mushroomdb suggest ./db) profiles your data and ranks candidate rules with estimated edge counts and rationale. Seeded sampling, so the same database always returns the same suggestions. No rule is ever applied automatically. docs/site/suggest.md
• An exception class per engine error. In Python, MushroomError(RuntimeError) is the base and eighteen subclasses hang off it — KeyNotFound, ReadOnly, Corrupt, CasConflict, RoleWriteDenied and the rest — each carrying .code, a stable snake_case string, and the variant's own fields as attributes. .code is the compatibility surface: classes may be added, a code is never respelled. Because the base subclasses RuntimeError, every except RuntimeError written against an earlier release keeps catching what it caught. A test asserts every Rust GraphError variant maps to a distinct class, so adding one without a class fails the build. db.roles() reads back what roles.json defines, so a sidecar can validate a role name at boot rather than on the first request.

CLI reference

Concurrency: every CLI write command, the hooks, and a running mushroomdb serve coordinate through one advisory LOCK file in the store directory, so they are safe to run against the same store at the same time. A writer that cannot get the lock within two seconds exits 3 with another mushroomdb process is writing; retry, having written nothing. Readers never take the lock and never wait; recall opens read-only (read_only: true) so an unattended hook can never delay a writer or fail because one is running. What the lock does not give you: cross-process transactions, and subscription events for a peer's writes — a commit absorbed by refresh() is visible on the next read but notifies nobody. Full model: docs/site/concurrency.md.

Full HTTP endpoint reference: docs/site/api.md.

Known limitations

Benchmarks

mushroomdb's own numbers. Each row names the committed file it comes from, and each file records its machine, date and command. Embedded: no figure includes a network round-trip.

Rows 1–7: benchmarks/results/mushroomdb-10k-0.7-ab.md, Apple M4 Pro, 1-minute load 3.4–3.8 on 12 cores. Each is the median of three runs of the harness, except the warm two-hop, which is the median of twenty runs of benchmarks/ab_driver.py; benchmarks/run.py reports the single pass only. The released 0.6.12 was measured alongside on the same day and is indistinguishable. That comparison is the driver's twenty runs a side, not the medians of three above: backfill 8.611 s for 0.6.12 against 8.584 s for this tree, ingest 0.991 s against 0.998 s, warm two-hop 195.3 µs against 192.1 µs. The first single run on this tree is mushroomdb-10k-0.7.md. Rows 8–9: dogfood/results/scale-100k.md, warm file cache, cold process; cold-cache was not measured.

Runs from 2026-08-21 and 2026-08-24 reported 2.85–3.51 s for the rule backfill. That was a different workload: before v0.2.0 a rule without an explicit max_edges stopped at a global cap, and since v0.2.0 it keeps the top 32 per source when declared through the Python binding, as the harness does, so the backfill evaluates every source. The earlier two-hop figure, 261.6 µs, was a warm median over 5.81M derived edges, a different edge set, and the earlier 784 ms ingest was a single shot (head-to-head-10k-v2.md, regression-v0.1-20260821.md, regression-v0.1.1-20260824.md).

Rule engine against hand-rolled maintenance (10,000 nodes, 1,000 specialty updates, drift 0 for all three): per-operation hand-written 64.93 min, batched hand-written 24.98 s, rule engine 17.58 s. Both hand-rolled variants were written by the engine team with full knowledge of retraction semantics — drift 0 is a property of that, not of hand-rolling in general. benchmarks/results/handrolled-vs-rules.md

Agent benchmarks measure the entity engine directly. benchmarks/agent-tasks/ runs real claude -p sessions against executable truth. The association suite (--suite association) asks twenty relationship questions of one generated world written three ways — as JSON files, as a single relational file, and as a mushroomdb store. The pre-registered gate passed on 0.6.12, in 20260928T195302Z, on a 2,000-entity world — the first run in which it has. The graph arm scored 1.000 on all twenty tasks and all sixty cells against the relational baseline's 0.990, at $0.0614 against $0.1127:

That run was on 0.6.12, whose default listing was 19 tools; 0.7's is 23 and about 15% larger, so the cost row is not 0.7's. The suite has not been re-run on 0.7.

180 cells, 0 timeouts, 0 errors, 0 dropped. Correctness had been amended on 2026-09-11 to pass on a tie, because the relational arm saturated at 1.00 and the original wording — exceed both baselines, interval excluding zero — was judged unpassable; this run passes the original wording. Every one of the six sub-1.0 cells belongs to a baseline, and each key they missed was named correctly by the graph arm in the same rep: classification.md.

The run before it (20260925T200950Z) failed on correctness by one key on one task. That key was a date-resolution defect in the engine, not a limit of the surface, and it is fixed in 0.6.12. The code suite (--suite code) is retired; its committed summaries stay as the record. docs/site/association-bench.md describes the suite and how to rebuild the world.

Architecture

graph-db/
├── crates/
│   ├── core-storage      # Packed adjacency topology + columnar property store + WAL + snapshots
│   ├── core-rules        # linking rules, per-rule indexes, incremental maintenance
│   ├── core-query        # pull-based interpreter; traversal ops + Cypher subset
│   ├── core-api          # the one public Rust interface; typed error enums
│   ├── code-extract      # tree-sitter symbol/import/call extraction; bytes in, facts out
│   ├── arrow-bridge      # results ↔ Arrow buffers
│   ├── server            # axum HTTP + WebSocket; serves UI
│   ├── cli               # mushroomdb binary
│   └── sim-harness       # DST: virtual clock, fault-injecting IO, seeded runner
├── ui/                   # TypeScript + Vite graph explorer
├── bindings/python/      # PyO3 / maturin
└── clients/typescript/   # HTTP + WebSocket client

Dependency rule (inward only): bindings/server/cli → core-api → {core-query, core-rules} → core-storage

Storage uses a dense-id WAL with per-commit fsync (configurable via FsyncPolicy), plus mmap-able V9 rkyv snapshots (13 sections: CSR topology, columnar properties, one shared string table for every string column, HNSW blobs, provenance, IVF state, per-node last-change index, and more — zero-copy, no heap allocation on open). V5–V8 stores are auto-migrated to V9 on GraphDb::open, keeping the original beside the new one as snapshot.bin.bak until the next clean open. The upgrade is one-way: an earlier binary refuses a V9 snapshot with snapshot: unsupported version 9 rather than reading it wrongly. Derived edges are not WAL-logged; they are restored directly from the mmap'd sections. See docs/format-stability.md for the format evolution contract.

Roadmap

What is not built yet:

Docs

• Quickstart · Rules · Cypher reference · HTTP + MCP API
• Concurrency · ingest-git as a data source
• Install, plugin and hooks · MCP tools · Association benchmark
• Time travel · Subscriptions · Views · Rule suggestions
• Masks and access control · Full-text search · Property indexes · Graph algorithms
• Durability and recovery · Running it as a service · Panic policy · Testing · Format stability
• Case study · The original design, 2026-08

Building from source, Docker, packaging, and the test gates are in CONTRIBUTING.md.

License

Copyright 2026 Matthew Sherlin.

Dual-licensed under MIT or Apache-2.0, at your option.