anatid

anatid is an embedded graph memory for AI agents, built on DuckDB and MIT licensed. The database is a single file with no server or daemon to run. pip install anatid gives an agent a memory that stores entities and the edges between them, records both what was true and what the agent believed at any past instant, and answers a query three ways at once (vector similarity, BM25 text, and 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.

Every retrieval structure in it is derived, not canonical. The full-text index, the graph CSR and the optional vector index are all built the same way: a versioned base generation plus a journal written in the same transaction as the row it describes, merged on every read before any tenant or time filter runs. A write is findable by the next read with nothing rebuilt, an index that is stale, damaged or absent costs latency rather than correctness, and every fallback says which of the eight reasons applies. The SQL path over the canonical tables is always the oracle.

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.

2-hop recall at 1,000,000 memories has a p50 of 2.88 ms on DuckDB against 7.35 ms on a tuned LadybugDB (the maintained MIT fork of Kuzu), a factor of 2.5, and the two engines return identical result id-lists. That measurement is why anatid is built on DuckDB. 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

Released on PyPI as anatid 0.2.0. To track main instead:

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

anatid runs on Python 3.10 through 3.13 and has one required dependency, duckdb>=1.5. CI runs the test suite on Linux and macOS across all four Python versions, including both integration suites; that is why [dev] installs openai-agents and mcp. 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, since 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()
    hits = db.recall("coffee", embedding=vec, seed_entity="Ada", k=3)   # findable already
    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 the output of running the block above. Nothing was rebuilt before that recall: the write was journalled inside its own transaction and the text arm merged it. db.maintain_indexes() folds the journal into new generations when you want the speed of a built index, and db.index_health() says whether that is due and why.

A longer, commented version covering recall_2hop, forget(hard=True) and stats() is in examples/quickstart.py. It needs no API key and finishes in under a second.

$ python examples/quickstart.py
anatid schema v4 on duckdb 1.5.5, tenant 1, expand path: sql
before rebuild: bm25 stale=False, rows merged from the journal=3
  findable with nothing built: ['Ada prefers dark roast coffee']
after rebuild : bm25 stale=False, index health=fresh

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=6 about_edges=2 supersedes_edges=1 audit_rows_deleted=1

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

Bo never appears in the query and has no edge to Ada. The graph arm reached that memory in two hops, Ada → Kestrel → ingest service.

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() 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, and taking its dry_run list first is the way to see what it will touch. Write verbs take now= and the temporal read verbs take as_of=, which keeps tests 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. The memory is ordinary tables, joinable against your Parquet and CSV in place.

Why DuckDB, with numbers

Phase 0 was a benchmark run before any of the library was written: 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 is the query shape agent memory hits hardest. Over 1,000 queries on a 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.
• LadybugDB's number is the fastest of six Cypher formulations across two thread settings. The naive formulation was 16x slower than the tuned one; reporting it would have flattered DuckDB.
• Where DuckDB loses: hybrid recall is 22% slower (16.4 ms against 20.0 ms p50; no engine in the run 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 covering every phase, p50/p95/p99, the mixed workload, concurrency, correctness, and nine 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 integration

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 supplies the missing three and 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)

• Writes are gated and reads are not. 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 approval policy is a callable. approve_low_risk() auto-approves small, ordinary writes and still stops for hard deletes. Unless you opt out explicitly, anatid_forget(hard=True) requires approval regardless, because a hard forget removes the row, its edges, its embedding and its provenance.
• Approval can happen later and 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, as a review queue rather than 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 rather than two round-trips to two different stores, and memories_written_here() reports what this conversation committed to memory.

MCP server

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

This exposes 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 the tool off with --no-sql-tool.

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

Limitations

Every item here is measured or documented in the source. Behavior that contradicts the docs and is not listed below is a bug; please report it.

• The default vector backend is still an exact scan. Anatid.open(vector_backend="duckdb_vss") opts in to an HNSW generation, which measured recall at k of 1.0000 (k=10) and 0.9982-0.9984 (k=50) against the exact oracle at 9,500 and 95,000 rows per tenant, and 2.2-2.8x the speed at 95,000. It is opt in because DuckDB documents HNSW persistence as experimental with write-ahead-log and crash-recovery caveats, because a persisted HNSW index silently loses its ef_search across a reopen (anatid reissues it per connection), and because below roughly 15,000 rows per tenant the exact scan is the faster of the two anyway. The 1M and 10M measurements the promotion criterion also names have not been taken. BRUTE_FORCE_CEILING = 100_000 is enforced since 0.1.1: on the exact backend recall(embedding=...) raises BruteForceCeilingError when the scan would cover more rows than that, unless you pass allow_slow=True. A usable generation lifts the ceiling, because the scan then covers only the journal.
• DuckDB's own full-text index is not incremental, and anatid builds incremental behaviour above it rather than exposing that. A write is journalled in its own transaction and merged into the next search, so .bm25_stale is False and the row is findable. What you still choose is when to pay for a rebuild: maintain_indexes() on a policy, or rebuild_fts_index() by hand. Two consequences. Merging costs read latency in proportion to the journal, not the corpus (measured: +2.3 ms at 500 journalled writes over a 100,000-document corpus). And with no generation published at all, a search scans the corpus exactly, which is refused above SCAN_CEILING = 100_000 documents per tenant; there 0.1.1's index answers if the file still has one, and the result says so.
• An index can be damaged in ways a read cannot afford to detect. Every read checks one cheap invariant per index and falls back to the oracle with HealthReason.damaged_base when it fails, but a base that is structurally consistent and wrong (postings lost from under a document map that still points at them, say) is caught by validate() during a rebuild, not by a read.
• One writing process per file. That is DuckDB's model, and the engine enforces it: a second read-write process cannot even open the file (IO Error: Could not set lock on file ...: Conflicting lock is held). Many threads inside that one process write concurrently and appends never conflict (0 errors in a 30 s, 6-thread benchmark with no retry logic), but anatid provides nothing for multi-process writes.
• Isolation is snapshot, not serializable. Two concurrent updates to the same row abort the second with a retryable ConflictError. anatid does not retry, because whether the write should be re-derived from a fresh read depends on the caller.
• Tenant isolation is file-per-tenant. DuckDB has no row-level or schema-level access control. A tenant_id column scopes queries; 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.
• DuckDB has no AS OF SYSTEM TIME. as_of() is a WHERE clause over valid_from/valid_to/tx_from/tx_to that anatid generates. It reaches back exactly as far as the rows still in the table, so a hard purge is gone from every as-of view too.
• The CSR still has sharp edges, though fewer than in 0.1. A generation numbers its own vertices, so dense entity ids are no longer required of you; build_csr()'s unnamed 0.1 snapshot still is, and still goes stale on any relate(). A generation is built in full rather than updated in place, so a large journal eventually costs more than the expansion saves (1.50 ms against 0.88 ms of pure SQL at about 550 journal rows on the spike graph), which is what MaintenancePolicy's ratio trigger prevents. The in-memory structure is not evicted by DuckDB's object cache, so memory grows with the number of resident generations. The C++ extension is still optional: without it the merge runs in SQL and returns the same rows.
• Maintenance is a call, not a thread. There is no background worker; maintain_indexes() runs when you run it.
• Pins are process-wide, not cross-process. A second process can only open the file read-only, so it cannot publish a generation, but it also cannot pin one against the writer process.
• No Cypher yet. Today the API is the verbs above plus SQL.
• This is v0.2. The API may still move, so pin the version.

Documentation

• docs/architecture.md: storage layout, the visibility predicate and the derived-index framework, the graph paths and how they stay MVCC-correct, the isolation contract, the temporal model, the recall pipeline.
• docs/design/derived-index-framework.md: the design the accelerators are built to, and what shipped in 0.2.0 against what was deferred.
• docs/benchmarks.md: Phase 0 method, every result, and what the benchmark does not tell you.
• docs/roadmap.md: v0.3 (Cypher subset, Graphiti/Cognee drivers, Node bindings), v0.5 (background maintenance, multi-process, graph algorithms), v1.0 (an owned ANN index, duckdb-wasm, format stability).
• 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.