khive
ai.khive/khive
Persistent memory, knowledge graph, task management and more for AI agents via MCP.
Documentation
khive
Vector search finds similar text. A knowledge graph finds structure: lineages, dependencies, contradictions, gaps. khive gives your research agent a typed, queryable graph that grows as it works. Read a paper and entities and edges fall out. Make a connection and it's traversable immediately. Come back next session and the graph remembers what you built.
There's no Neo4j to run and no separate SPARQL endpoint to deploy. It's SQLite on disk, MCP over
stdio, and cargo test finishes in 4 seconds.
khive also runs hosted, as memory and a knowledge graph your agents reach over MCP: https://khive.ai/?utm_source=github&utm_campaign=khive-readme
What you get
| Capability | How |
|---|---|
| 162 verbs, 14 packs | KG, GTD, memory, brain, comm, schedule, knowledge, session, tool, exec, git, code, workspace, blob: all load by default |
| Typed entities | 9 closed kinds: concept, document, dataset, project, person, org, artifact, service, resource |
| Typed edges | 20 closed relations in 10 categories (structure, derivation, provenance, temporal, dependency, impl, lateral, annotation, epistemic, ownership) |
| Typed notes | 5 closed kinds: observation, insight, question, decision, reference |
| Hybrid retrieval | FTS5 + vector RRF with embedding rerank; shipped BM25, HNSW, Vamana, and fusion crates for pack-specific retrieval paths |
| Graph traversal | BFS with depth/direction/relation filters, bidirectional shortest path |
| GQL + SPARQL queries | Parse to SQL, run against the same SQLite backend |
| Salience-weighted notes | Notes carry salience scores; search ranks by semantic relevance × salience |
| Cross-substrate links | Notes annotate entities (and vice versa) via the same edge system |
| Soft delete + supersede | History-preserving: old records stay, newer ones supersede via graph edges |
| Namespace attribution | Every record is stamped with a namespace and all records share one store by default. Operators can supply a custom Gate for namespace-scoped enforcement |
The three substrates
Everything in khive is one of three things:
| Substrate | What it is | Mutability | Example |
|---|---|---|---|
| Entity | A graph node with typed edges | Mutable + soft-delete | LoRA (concept), arxiv:2106.09685 (document) |
| Note | A temporal observation about the world | Mutable + soft-delete | "FlashAttention gains scale with seq len, not batch" |
| Event | An audit log entry | Immutable | create(kind="entity", ...) was called at T |
Entities are things. Notes are what you think about things. Events are what happened.
The MCP verb surface
One MCP tool: request (ADR-016 + ADR-027). Every verb is a parsed op inside it.
request(ops="verb(arg=value, arg=value)") # single op
request(ops="[v1(...), v2(...), v3(...)]") # parallel batch (max 100)
request(ops="[{\"tool\":\"v1\",\"args\":{...}}, ...]") # equivalent JSON form
All 14 packs load by default, giving 162 verbs out of the box (updated from the
current handler declarations; verify again with request(ops="verbs()")
before editing this table):
| Pack | Prefix | Verbs | What it does |
|---|---|---|---|
| kg | (bare) | 31 | Entities, edges, notes, graph queries, reference resolution, caller identity, WAL diagnostics |
| gtd | gtd. | 8 | Task lifecycle, per-assignee counts, timestamp census, and explicit historical repair |
| memory | memory. | 6 | Salience-weighted remember / decay-ranked recall |
| brain | brain. | 17 | Bayesian user profiles + feedback loop |
| comm | comm. | 15 | Threaded messaging with sender-side internal delivery confirmation |
| schedule | schedule. | 4 | Reminders and scheduled verb execution |
| knowledge | knowledge. | 21 | Atom-based KB with embedding rerank search |
| session | session. | 7 | Session records, dependency-gated mirror search, store stats, and compaction (store/list/resume/export/search/stats/vacuum) |
| tool | tool. | 14 | Capability registry, discovery by capability, and use policy: check/request/grant (ADR-180) |
| exec | exec. | 9 | One declared command in a sandbox over a materialized tree, receipts for every run (ADR-181) |
| git | git. | 17 | Provenance ingest, branch/commit/update_ref/push writes (ADR-108), dev-loop verbs: checkout, diff, gates, receipts, reconcile, status, log, init, PR open/review/merge (ADR-182) |
| code | code. | 4 | Source ingest and registered-map list, search, dependency reads (ADR-085, ADR-139) |
| workspace | (none) | 0 | Adds the workspace entity kind + contains endpoint rules to git/gtd/session notes (#873) |
| blob | blob. | 9 | Content-addressed put/get/stat, staged begin/put_part/commit/abort, and opt-in blob.import/blob.export file transfers over BlobStore (ADR-111, ADR-173) |
create, list, search take kind=entity|note (or kind=edge for list).
get, update, delete, merge are UUID-only: they auto-detect the record type.
Agents reach khive via MCP stdio, using Python, TypeScript, Rust, or any MCP-compatible client, so there is no language SDK to learn.
What typed edges add beyond vector similarity
A vector index returns "these two texts are close." It has no idea that document B introduced concept A, that concept C is a variant of concept A, or that concept D was superseded by concept E last month. Cosine distance carries no direction and no type. It can't tell you what a relationship is, only that something is nearby.
khive's graph carries both signals. Every edge is one of 20 closed relations across 10
categories: structure (contains, part_of, instance_of, links_to, located_in), derivation (extends,
variant_of, introduced_by, supersedes), provenance (derived_from), temporal
(precedes), dependency (depends_on, enables), implementation (implements), lateral
(competes_with, composed_with), annotation (annotates), epistemic (supports,
refutes), and ownership (owns, owner → owned).
The following example, run against a scratch database, shows the difference in practice
(full transcript in demos/research-ingest.md):
request(ops="create(kind=\"entity\", entity_kind=\"concept\", name=\"FlashAttention\")")
request(ops="create(kind=\"entity\", entity_kind=\"concept\", name=\"FlashAttention-2\", description=\"Improved parallelism and work partitioning over FlashAttention\")")
request(ops="link(source_id=\"<fa2_id>\", target_id=\"<fa_id>\", relation=\"extends\")")
request(ops="traverse(roots=[\"<fa_id>\"], max_depth=2)")
traverse returns the lineage directly: FlashAttention-2 reaches FlashAttention across one
extends edge, with the edge id and relation name attached to the path. A vector search over
the same two entities returns a similarity score between two chunks of text, with no way to tell
which one came first or which direction the relationship runs.
khive runs both signals together. search combines FTS5 and vector similarity through RRF
fusion to find candidates. traverse, neighbors, and GQL/SPARQL walk the typed edges to show
how those candidates actually relate. Similarity surfaces what's nearby in meaning. The graph
records what's connected, in which direction, and why.
Architecture
┌──────────────────────────────────────────────────────────────┐
│ kkernel mcp: stdio MCP server (the `kkernel` binary) │
│ khived: persistent daemon (ADR-049): warm runtime │
│ auto-spawned on first request │
│ 1 tool: `request` (ADR-016 + ADR-027): parses DSL, │
│ dispatches each op through the VerbRegistry │
└──────────────────────────────────────────────────────────────┘
↕ VerbRegistry dispatch
┌──────────────────────────────────────────────────────────────┐
│ khive-pack-kg: KG vocabulary + 31 public verbs │
│ khive-pack-gtd: task lifecycle (8 verbs) │
│ khive-pack-memory: salience + decay recall (6 verbs) │
│ khive-pack-brain: Bayesian profiles (17 verbs) │
│ khive-pack-comm: threaded messaging (15 verbs) │
│ khive-pack-schedule: reminders + scheduled ops (4 verbs) │
│ khive-pack-knowledge: atom KB + embedding rerank (21 verbs) │
│ khive-pack-session: session records, search, stats, vacuum (7 verbs) │
│ khive-pack-git: provenance ingest + writes (17 verbs) │
│ khive-pack-code: source ingest (L1 manifest + L1.5 │
│ import-scan) + map reads (4 verbs) │
│ khive-pack-workspace: workspace entity + contains endpoint │
│ rules (0 verbs) │
│ khive-pack-blob: content-addressed object storage │
│ (9 verbs) │
│ khive-pack-tool: tool/skill/plugin registry, capability │
│ discovery, use policy (14 verbs) │
│ khive-pack-exec: sandboxed run over a materialized tree │
│ (9 verbs) │
└──────────────────────────────────────────────────────────────┘
↕ in-process
┌──────────────────────────────────────────────────────────────┐
│ khive-runtime, khive-request, khive-query, khive-db, │
│ khive-storage, khive-score, khive-types │
└──────────────────────────────────────────────────────────────┘
Embedded SQLite storage with FTS5 trigram text search and sqlite-vec vector search. Retrieval ships in-process BM25, HNSW, Vamana, and fusion crates for hybrid and pack-specific search paths. One binary, one DB file, no external services to run.
The khived daemon (ADR-049) keeps the runtime warm between MCP sessions: the embedding model stays loaded, SQLite connections stay open, and pack registries stay initialized. It auto-spawns on first request and persists in the background, eliminating cold-start overhead on reconnect.
HTTP gateway and visual frontend are planned for future releases. The kkernel admin CLI
(migrations, reindexing, data import/export, diagnostics) ships today, documented in
docs/operations.md.
Performance
Knowledge search runs on an in-process Vamana ANN index (khive-vamana). On the standard
SIFT-1M benchmark, the index returns recall@10 of 0.95 at a p50 query latency of 171µs over
1,000,000 vectors, measured on a single laptop (macos-arm64, commit eb6696c). Tail latency
stays under 250µs and the index scales sublinearly as the corpus grows from 100K to 1M vectors:
| Vectors | Recall@10 | p50 | p95 | p99 | Build |
|---|---|---|---|---|---|
| 100K | 0.9504 | 71µs | 93µs | 102µs | 11.1s |
| 316K | 0.9523 | 130µs | 177µs | 200µs | 54.6s |
| 1M | 0.9521 | 171µs | 216µs | 234µs | 320.8s |
Query latency grows about 2.4x while the corpus grows 10x, so the search path is sublinear in corpus size over this range. Index build is a one-time cost paid at construction; it grows super-linearly here (about 29x build time for the 10x corpus).
The speedups over exhaustive search implied by these latencies (89x at 100K, 153x at 316K, 341x
at 1M) are computed against a back-derived brute-force baseline rather than a directly measured
one (see #167); treat them as indicative, not
as a measured headline figure. The benchmark harness lives in perf/; raw data is in
perf/ledger.csv. New latency and throughput claims used as decision evidence
must follow the cache-state and warm-up protocol.
Crates
| Crate | Purpose |
|---|---|
khive-types | Domain types, Pack trait, closed enums |
khive-score | Deterministic i64 fixed-point scoring |
khive-storage | Trait-only capability surface (zero implementations) |
khive-db | SQLite backend: entity/note/edge tables, FTS5 TextSearch, current sqlite-vec VectorStore compatibility |
khive-retrieval | Hybrid retrieval primitives |
khive-fusion | RRF, weighted, union, vector-only, and keyword-only fusion strategies |
khive-bm25 | BM25 keyword index |
khive-hnsw | HNSW vector index |
khive-vamana | Vamana ANN index used by knowledge search |
khive-query | SPARQL / GQL → SQL compiler |
khive-runtime | Service API + VerbRegistry + PackRuntime trait |
khive-request | Request DSL parser (function-call, JSON; pipe / LNDL planned). Transport-agnostic AST. |
khive-pack-kg | KG pack: vocabulary, verb handlers, kind validation |
khive-pack-gtd | GTD pack: task lifecycle over the notes substrate |
khive-pack-memory | Memory pack: salience-weighted remember/recall with decay |
khive-pack-brain | Brain pack: Bayesian user profiles, feedback, resolution |
khive-pack-comm | Comm pack: threaded messaging with inbox |
khive-pack-schedule | Schedule pack: reminders and scheduled verb execution |
khive-pack-knowledge | Knowledge pack: atom-based KB with embedding rerank search |
khive-pack-session | Session pack: session record persistence |
khive-pack-git | Git pack: provenance ingest and git write verbs |
khive-pack-code | Code pack: L1 manifest + L1.5 import-scan source ingestion |
khive-pack-workspace | Workspace pack: workspace entity kind + contains endpoint rules |
khive-pack-tool | Tool pack: capability registry, discovery, use policy (ADR-180) |
khive-pack-exec | Exec pack: sandboxed run of a registered tool over a materialized tree (ADR-181) |
khive-pack-blob | Blob pack: content-addressed object storage |
khive-mcp | MCP server library: single request tool dispatching through the VerbRegistry (served by kkernel mcp) |
kkernel | The single shipped binary: kkernel mcp serves MCP; admin subcommands (exec, reindex, db, …) |
Dependency direction (storage stack): types → score → storage → db → query → runtime → packs → mcp.
Side input: request → mcp (the DSL parser is consumed only at the MCP dispatch boundary;
packs do not depend on it).
Storage is trait-only; backends (SQLite today, Postgres tomorrow) implement the traits without
touching consumers.
Quick start
Use khive cloud for the hosted service.
Run it locally from source
For local development, install Rust via rustup if needed, then install
kkernel from a source checkout:
git clone https://github.com/ohdearquant/khive.git && cd khive
cargo install --path crates/kkernel --locked
kkernel is the single shipped binary; kkernel mcp serves the MCP request surface.
Add to your MCP config (.mcp.json in your project, or ~/.claude/mcp.json for
global):
{ "mcpServers": { "khive": { "command": "kkernel", "args": ["mcp"] } } }
Verify:
kkernel --version # confirms the binary and version you just installed
All 14 packs load by default, a background daemon auto-spawns to keep the runtime warm, and any
MCP client discovers the request tool with the full 162-verb catalog.
Usage
The agent expresses verbs as DSL ops inside the single request tool:
request(ops="create(kind=\"entity\", entity_kind=\"concept\", name=\"LoRA\")")
request(ops="search(kind=\"entity\", query=\"parameter efficient fine-tuning\")")
request(ops="link(source_id=\"<uuid>\", target_id=\"<uuid>\", relation=\"variant_of\")")
# Batch multiple ops in one call:
request(ops="[create(...), create(...), link(...)]")
Claude Desktop
Add the same server entry to claude_desktop_config.json:
{ "mcpServers": { "khive": { "command": "kkernel", "args": ["mcp"] } } }
See docs/guide/getting-started.md for the config file location on each platform.
Claude Code plugin (skills + agents)
For guided workflows, install the marketplace plugin:
/plugin marketplace add ohdearquant/khive
/plugin install khive
This adds one pattern skill per pack (kg, gtd, memory, brain, comm, schedule,
knowledge), four kg workflow skills, and six kg stewardship agents:
| Skill | What it does |
|---|---|
/khive:digest | Ingest material into the graph: extract entities, link, verify |
/khive:gap | Survey the graph's structural gaps: read-only frontier ranking |
/khive:expand | Grow the graph to close one strategic gap, with create caps |
/khive:polish | Audit and fix: orphans, low-degree nodes, duplicates |
The agents (khive:digester, khive:gap-analyst, khive:expander, khive:polisher,
khive:librarian, khive:researcher) and the pattern skills are described in
marketplace/khive/README.md.
Configuration
khive mcp # Default: ~/.khive/khive.db
khive mcp --db /path/to/my.db # Custom DB path
khive mcp --db :memory: # Ephemeral (testing)
khive mcp --namespace my-project # Default namespace (default: "local")
khive mcp --no-embed # Disable local embedding model
khive mcp --log debug # Log level (default: warn)
Environment variables: KHIVE_DB, KHIVE_NAMESPACE, KHIVE_NO_EMBED, KHIVE_LOG.
For config file discovery order, the [[backends]] model, and how --db interacts
with a declared backend topology, see docs/configuration.md.
For the kkernel admin CLI (migrations, reindexing, data import/export, diagnostics), see
docs/operations.md.
Development
cd crates && cargo test --workspace
make ci # Full CI: fmt, clippy, test, build
Prerequisites: Rust 1.95+ (workspace MSRV) (via rustup), Deno 2.x (for the TypeScript CLI layer, optional)
- Node.js 20+ and pnpm (for frontend, optional)
Demos
Runnable, copy-pasteable transcripts against a scratch database. See demos/:
demos/research-ingest.md: create entities, link them, search, and traverse the graphdemos/gtd-memory.md: task lifecycle and salience-weighted memory recall
Docs: ohdearquant.github.io/khive (agents: fetch
/llms.txt). Questions and discussion:
Discord.
Contributing
- Feature branches + PRs. Never push directly to main.
make cimust pass (fmt, clippy, test, no-default-features check, release build).- Conventional commits:
feat(types): add NoteKind taxonomy. - Schema/interface changes need a design doc. Propose it in the PR or as an issue.
- See CLAUDE.md for the developer guide, AGENTS.md for agent usage.
Status
162 verbs across 14 packs, 9 entity kinds, 20 edge relations, daemon warm startup (ADR-049), knowledge search with embedding rerank, Bayesian brain profiles, threaded messaging, scheduled verb execution. Ready for use with Claude Code and any MCP-compatible agent.
License
Apache 2.0. See LICENSE.
No installation packages available.