cairn

agent wayfinding. temporal knowledge store in one sqlite file. no neo4j, no cloud, no lock-in.

landing - github

what

cairn is a new category. not "memory" or "knowledge graph." agent wayfinding - agents that can find their way back to what they know.

a cairn is a pile of stones hikers stack to mark a path. agents leave cairns to mark what they learned and where they've been. each fact is a stone. the pile is the path.

MCP gave agents tools. cairn gives them a path.

why

every agent memory system today is either a walled garden (mem0, zep cloud) or a heavy dependency (neo4j, postgres). none of them talk to each other. your memory is locked into whichever platform you picked first. cairn is the opposite: one binary, one file, zero dependencies, works with any agent that speaks MCP.

	mem0	zep	neo4j	cairn
setup	pip + api key	docker + api key	docker + cypher	one binary
dependencies	python, redis	docker, postgres	jvm, 2gb+ ram	sqlite
protocol	rest api	rest api	cypher/bolt	MCP
temporal queries	no	limited	no	yes
forgetting	no	no	no	yes
federated sync	no	no	no	yes
works offline	no	no	yes	yes

how it works

cairn stores facts as temporal triples: subject, predicate, object, with validity windows.

tamish --uses_os--> macos    (valid: 2024-01-01 to 2025-07-15)
tamish --uses_os--> linux     (valid: 2025-07-15 to now)

when a new fact contradicts an old one, the old fact is closed (not deleted). you can query the past: "what did we know about tamish in march?"

five things cairn does that nobody else does together

1. bi-temporal tracking without a graph database. every fact carries two timestamps: when it was true in the world, and when the system learned it. contradicted facts get closed, not deleted. all in sqlite.

2. extraction without burning LLM calls. the agent calls remember() as an MCP tool during its normal response. no separate extraction pipeline. no extra API calls. a pattern-based post-turn hook catches 80% of facts for free.

3. forgetting as a first-class operation. facts decay by confidence, access frequency, and age. stale facts get tombstoned (soft delete with reason). forgetting is auditable. memory gets better over time, not just bigger.

4. federated sync. your phone agent and laptop agent share a brain. peer-to-peer sync via vector clocks. conflict resolution is deterministic. no server required.

5. query by meaning, not by query language. the agent says recall("what do you know about tamish's setup") and gets back a subgraph. no cypher, no SQL. semantic similarity finds entry points, graph traversal follows relationships, ranking returns what matters.

architecture

cairn/
  crates/
    store/        temporal sqlite engine, validity windows, conflict resolution
    traverse/     graph traversal, depth-limited
    forget/       decay scoring, garbage collection, tombstones
    query/        semantic entry points + traversal + ranking
    extract/      pattern-based fact extraction (no LLM)
    sync/         vector clocks, peer sync, conflict resolution
    mcp/          MCP server exposing remember/recall/forget/export
  bin/
    cairn/        CLI + MCP server entry point

one sqlite file. one binary. zero external services.

install

cargo install cairn-memory

or build from source:

git clone https://github.com/valtors/cairn
cd cairn
cargo build --release
cp target/release/cairn /usr/local/bin/

usage

# run as MCP server (any MCP-compatible agent connects)
cairn serve

# or use directly
cairn remember --subject tamish --predicate uses_os --object linux
cairn recall "what os does tamish use"
cairn forget --older-than 30d
cairn export > my-memory.json
cairn import < my-memory.json

with claude desktop

{
  "mcpServers": {
    "cairn": {
      "command": "cairn",
      "args": ["serve"]
    }
  }
}

tests

63 tests across 6 crates. all pass.

cargo test --workspace

tech

rust sqlite serde rusqlite MCP temporal triples vector clocks single binary zero runtime deps

contributing

see CONTRIBUTING.md. we welcome contributions of all kinds - bug fixes, new patterns, transport support, docs.

good first issues are labeled good first issue.

license

MIT. strictly open source. no cloud tier, no enterprise plan, no proprietary fork.