anatid

Embedded graph memory for AI agents, built on DuckDB. MIT licensed.

One file on disk. No server, no daemon, no cluster. pip install anatid and your agent has a memory that is a graph (entities and the edges between them), bitemporal (what was true, and what you believed, at any past instant), and searchable three ways at once (vector similarity, BM25 text, graph traversal — fused into one ranked list). Writes can be routed through the OpenAI Agents SDK's human-in-the-loop approval flow, so an agent proposes a change to its memory and a person decides whether it lands.

anatid exists because Kuzu was archived on 2025-10-10. Graphiti deprecated its Kuzu driver, Mem0 removed open-source graph memory in v2.0.0, and Cognee is migrating away. That left a lot of people with an embedded graph memory and nowhere to go. anatid is somewhere to go, and it is built on an engine with a foundation behind it.

The engine choice was not a preference. It was a measurement: 2-hop recall at 1,000,000 memories runs 2.6x faster on DuckDB than on a tuned LadybugDB (the maintained MIT fork of Kuzu), returning byte-identical result lists. The numbers, the method, and the caveats are in docs/benchmarks.md.

Install

pip install anatid                    # just duckdb
pip install "anatid[agents]"          # + the OpenAI Agents SDK integration
pip install "anatid[mcp]"             # + the MCP server

The version in this repository is 0.1.0 and is not on PyPI yet. Until it is:

pip install "git+https://github.com/thedatasense/anatid"

Python 3.10-3.13, one required dependency (duckdb>=1.5). CI runs the test suite — including both integration suites, which is why [dev] installs openai-agents and mcp — on Linux and macOS across all four Python versions. The tests that load the 100k-row spike dataset and the ones that need the compiled C++ extension skip in CI, because neither is in the repository; they are run locally before a release. Windows should work (DuckDB supports it) but is not tested.

Quickstart

from anatid import Anatid, utcnow

vec = [0.0] * 63 + [1.0]                                        # your embedding model's output

with Anatid.open("agent.anatid", tenant=1, embedding_dim=64) as db:
    db.relate("Ada", "Kestrel", rel_kind="leads")               # an entity -> entity edge
    m = db.remember("Ada prefers dark roast coffee",            # a fact, filed under 2 entities
                    entities=["Ada", "coffee"], kind="preference",
                    embedding=vec, writer="agent-1",
                    episode="Standup 2026-03-01: Ada takes it dark roast.")  # raw evidence first
    t0 = utcnow()
    db.rebuild_fts_index()                                      # BM25 is not incremental: you say when

    hits = db.recall("coffee", embedding=vec, seed_entity="Ada", k=3)
    print(hits[0].content, hits[0].sources, "| bm25_stale:", hits.bm25_stale)

    new = db.supersede(m.memory_id, "Ada switched to decaf")    # closes the old row, keeps it
    print("old still current?", db.get(m.memory_id).is_current) # False -- history is intact
    print("at t0:", [x.content for x in db.as_of(t0).recall_2hop("Ada")])
    print("evidence:", db.provenance(new.memory_id).source_text)

Ada prefers dark roast coffee ('vector', 'text', 'graph') | bm25_stale: False
old still current? False
at t0: ['Ada prefers dark roast coffee']
evidence: Standup 2026-03-01: Ada takes it dark roast.

That is real output from running the block above. A longer, commented version covering recall_2hop, forget(hard=True) and stats() is in examples/quickstart.py — it needs no API key and runs in about a second:

$ python examples/quickstart.py
anatid schema v2 on duckdb 1.5.5, tenant 1, expand path: sql
before rebuild: bm25 stale=True, rows waiting=3

recall(query + embedding + seed): arms=('vector', 'text', 'graph') stale=False
  [1] 0.0487 'Ada prefers dark roast coffee' via vector+text+graph about=['Ada', 'coffee']
  [2] 0.0325 'The ingest service is maintained by Bo' via vector+graph about=['ingest service', 'Bo']
  [3] 0.0320 'Ada leads Project Kestrel' via vector+graph about=['Ada', 'Kestrel']

recall_2hop('Ada'):
  'The ingest service is maintained by Bo'
  'Ada leads Project Kestrel'
  'Ada prefers dark roast coffee'

supersede: old is_current=False valid_to=2026-03-31 09:00:00 -> new 'Ada switched to decaf'

as_of(day 1)  : ['The ingest service is maintained by Bo', 'Ada leads Project Kestrel', 'Ada prefers dark roast coffee']
current       : ['Ada switched to decaf', 'The ingest service is maintained by Bo', 'Ada leads Project Kestrel']

provenance(depth=1, writers=['agent-2', 'agent-1']):
  current  'Ada switched to decaf' (by agent-2)
  closed   'Ada prefers dark roast coffee' (by agent-1)
  source: 'Standup 2026-03-01: Ada is leading Project Kestrel; she take'...

forget(hard=True): rows_removed=5 about_edges=2 supersedes_edges=1 audit_rows_deleted=1

stats: memories=3 current=2 entities=5 about=6 relates=2

Note the third line of that output. Bo is never mentioned in the query and is not connected to Ada directly — the graph arm reached it in two hops (Ada → Kestrel → ingest service), which is the thing a vector store cannot do for you.

The verbs

Each write verb is exactly one DuckDB transaction. Reads (recall, recall_2hop, context, get, provenance, stats) run their statements outside an explicit transaction, so a concurrent commit can land between a recall's arms and its hydration step — wrap the call in db.transaction() yourself if you need one snapshot. prune is a query plus one transaction per memory it forgets, so a failure part-way leaves the earlier deletions committed; take its dry_run list first. Every verb takes now=/as_of= so tests are deterministic. There are function forms too (from anatid.verbs import remember), and db.connection hands you the raw DuckDB cursor whenever you want to write SQL — your memory is just tables, joinable against your Parquet and CSV in place.

Why DuckDB, with numbers

Phase 0 was a benchmark, not a design document: 1,000,000 memories, 2.3M edges, 10 tenants, four engines, the same operations with identical semantics, checked against a pure-Python oracle.

2-hop recall (the query agent memory hits hardest), 1,000 queries, single thread:

• The kill criterion was "abandon DuckDB if it is more than 5x slower". It came in at 0.39x (SQL) and 0.28x (extension). At p95 it is 0.12x and 0.11x.
• All three engines returned identical result id-lists on 1,000 oracle-checked queries and on 200 post-write verify queries. Fast and wrong is not interesting.
• LadybugDB's number is its best of six Cypher formulations across two thread settings. The naive formulation was 17x slower; reporting that one would have been dishonest.
• Where DuckDB loses: hybrid recall is ~20% slower (16.4 ms vs 20.0 ms p50 — no engine had an ANN index, so this is a scan-speed comparison), and concurrent readers cost DuckDB writers real throughput (397 W1/s with writers alone, 152-189 W1/s with 2 readers added). LadybugDB with enable_multi_writes=True commits more writes per second than DuckDB does.

Full tables — every phase, p50/p95/p99, mixed workload, concurrency, correctness, and nine explicit limitations of the benchmark itself — are in docs/benchmarks.md. The raw JSON with per-operation latency arrays is in spike/results/.

OpenAI Agents SDK: memory the agent proposes and a human approves

The OpenAI Agents SDK already has everything needed for human-in-the-loop: needs_approval=True on a function_tool, RunResult.interruptions, a serializable RunState, state.approve() / state.reject(). It also has a Session protocol for conversation history, with SQLite, SQLAlchemy and Redis backends.

What it does not have is a DuckDB session, graph memory, or approval-gated memory writes. As far as we can establish, no open-source project combines all four of the Agents SDK, DuckDB, a graph store, and human approval on memory writes. anatid is the missing piece, and it rebuilds none of the SDK's machinery:

from agents import Agent, Runner
from anatid import Anatid
from anatid.integrations.openai_agents import AnatidSession, create_memory_tools

db = Anatid.open("agent.anatid", tenant=1)
session = AnatidSession("conv-1", db)                # conversation history, same file as the graph
tools = create_memory_tools(db, session=session)     # 3 read tools + 3 write tools

agent = Agent(name="assistant", tools=tools)
result = await Runner.run(agent, "Ada switched to decaf, remember that", session=session)

while result.interruptions:                          # writes stop here; reads never do
    state = result.to_state()
    for item in result.interruptions:
        print(item.tool_name, item.raw_item.arguments)   # "anatid_remember" {"content": ...}
        state.approve(item)                              # or state.reject(item)
    result = await Runner.run(agent, state, session=session)

• Reads are not gated; writes are. anatid_remember, anatid_supersede and anatid_forget carry needs_approval; anatid_recall, anatid_context and anatid_provenance do not. Nothing touches the database until someone approves.
• The policy is a callable. approve_low_risk() auto-approves small, ordinary writes and still stops for hard deletes. anatid_forget(hard=True) requires approval regardless unless you opt out explicitly — a hard forget removes the row, its edges, its embedding and its provenance, and that is not a decision to delegate to a model.
• Approval can happen later, elsewhere. RunStateStore(db) parks the SDK's serialized RunState in the same anatid file, so an interrupted run can be reviewed and resumed minutes or days later by another process — a review queue, not a blocking prompt.
• History and knowledge are joinable, because AnatidSession writes turns into a table in the same DuckDB file as the memory graph. await session.entities_mentioned() is one SQL join against entities, not two round-trips to two different stores; memories_written_here() tells you what this conversation actually committed to memory.

MCP server

pip install "anatid[mcp]"
anatid-mcp --db memory.anatid        # stdio; point Claude Desktop, Claude Code or Cursor at it

The memory verbs over the Model Context Protocol, so any MCP client gets persistent, bitemporal, graph-shaped memory: remember, relate, supersede, reinforce, forget, prune and rebuild_fts_index on the write side, recall, context, get, provenance and stats on the read side. (Those are the MCP tool names; the anatid_-prefixed names belong to the Agents SDK integration above.) --read-only registers the read tools only.

There is also one deliberate escape hatch: a sql tool for the questions the verbs do not answer ("how many memories per kind?", "show me the audit trail"). It is read-only, and enforced in three layers by DuckDB rather than by a regex over the query text: DuckDB's own statement classifier (only SELECT/EXPLAIN, and every statement in the text must pass), a scan of DuckDB's parse tree for file-reading functions and for base-table names that are not plain identifiers (DuckDB's replacement scan makes SELECT * FROM '/etc/passwd.csv' an ordinary SELECT), and execution inside BEGIN TRANSACTION READ ONLY on a private cursor that is always rolled back. DuckDB will not give a second read-only connection to a file the process already holds, so the read-only transaction is the mechanism. PRAGMA create_fts_index(...), which expands into DDL at bind time, is rejected on what it really is. Turn it off with --no-sql-tool.

from anatid.integrations.mcp import build_server if you want to embed the server in your own process.

What this is not, yet

Every item here is measured or documented in the source, not a guess. If you hit one of these, you were warned; if you hit something that is not here, that is a bug and we want the report.

• No ANN index. The vector arm is a brute-force array_cosine_similarity scan, because DuckDB ships no ANN index. The cost is linear in one tenant's row count: measured at 64 dims on the spike hardware, 2.0 ms p50 with 10k memories in the tenant, 8.6 ms at 100k (an independent run of the same measurement got 11.4 ms) and 23.3 ms at 1M. BRUTE_FORCE_CEILING = 100_000 is documented and not enforced — you are already paying ~9-11 ms per recall at that ceiling, and past it this is the wrong tool. An owned ANN index is the headline item of v1.0.
• The full-text index is not incremental. DuckDB's fts index does not see rows inserted after it was built. anatid does not paper over this: rebuild_fts_index() is explicit, fts_status() tells you how stale you are, and every recall() result carries .bm25_stale and .pending_fts_rows (with on_stale_fts="error" if you would rather raise). The staleness window is the gap between your rebuilds, and it is yours to choose.
• One writing process per file. That is DuckDB's model, and it is enforced by the engine: a second read-write process cannot even open the file (IO Error: Could not set lock on file ...: Conflicting lock is held). Many threads in that process write concurrently and appends never conflict (0 errors in a 30 s, 6-thread benchmark with no retry logic), but multi-process writes are not something anatid provides.
• Snapshot isolation, not serializable. Two concurrent updates to the same row abort the second with a retryable ConflictError. Retrying is your call, because only you know whether re-reading first changes the write.
• Tenant isolation is file-per-tenant. DuckDB has no row-level or schema-level access control. A tenant_id column is scoping; the real boundary is one file per tenant via DatabasePool, enforced by the filesystem. Raw SQL through db.connection sees every tenant in the file, and the docstrings say so.
• Time travel is our filter, not the engine's. DuckDB has no AS OF SYSTEM TIME. as_of() is a WHERE clause over valid_from/valid_to/tx_from/tx_to. It reaches back exactly as far as the rows still in the table — a hard purge is gone from every as-of view too, which is the point of a hard purge.
• The CSR extension has sharp edges. It needs dense per-tenant entity ids (anatid's default 63-bit time-ordered ids are not dense), it is rebuilt in full rather than incrementally, and any relate() marks it stale — at which point recall silently falls back to the SQL path, which returns identical rows. It is an accelerator, off by default.
• No Cypher yet. v0.2. Today the API is the verbs above plus SQL.
• v0.1. The API may still move. Pin the version.

Documentation

• docs/architecture.md — storage layout, the derived CSR and how it stays MVCC-correct, the isolation contract, the temporal model, the recall pipeline.
• docs/benchmarks.md — Phase 0 method, every result, and what the benchmark does not tell you.
• docs/roadmap.md — v0.2 (Cypher subset, Graphiti/Cognee drivers, Node bindings), v0.5 (production operation), v1.0 (ANN index, persistent CSR, duckdb-wasm).
• CONTRIBUTING.md — how to build it, what we care about in a change, and the third-party notices.
• spike/ — the Phase 0 evidence, kept read-only.

License

MIT. Copyright (c) 2026 anatid contributors. Code adapted from DuckDB (MIT) or from Kuzu / LadybugDB (MIT, Copyright 2022-2025 Kùzu Inc.) carries its original notice alongside ours; see CONTRIBUTING.md.