attestdb 0.1.32

The database for knowledge you can verify

Attest

Licensed under the Business Source License 1.1. Free to use, modify, and self-host. Cannot be offered as a competing managed service. Converts to Apache 2.0 after 4 years.

The database for agents. Persistent memory that compounds across sessions -- every fact sourced, every claim confidence-weighted, every failure a data point that prunes the search tree for everyone. pip install, single file, no server. 1.3M claims/sec ingestion, 8us entity query.

Install

pip install attestdb

Quick Start

import attestdb

db = attestdb.open("my_knowledge.db")

# Ingest a claim with provenance
db.ingest(
    subject=("BRCA1", "gene"),
    predicate=("associated_with", "relates_to"),
    object=("Breast Cancer", "disease"),
    provenance={"source_type": "literature_extraction", "source_id": "PMID:12345"},
    confidence=0.9,
)

# Query the knowledge graph
frame = db.query("BRCA1", depth=2)
print(frame.focal_entity.name)
for rel in frame.direct_relationships:
    print(f"  --[{rel.predicate}]--> {rel.target.name} (conf={rel.confidence:.2f})")

db.close()

AI Agent Memory

Attest gives AI coding agents persistent, cross-session memory. Install the MCP server and your agent learns from every session -- warnings, patterns, dead ends, and fixes are recalled automatically.

pip install attestdb[mcp]
attest-mcp install          # Auto-detect Claude Code, Cursor, Windsurf, Codex

What it does:

• Recall hook -- injects prior knowledge at session start (warnings, patterns, next steps from last session)
• Pre-edit check -- surfaces known bugs/warnings before you edit a file
• Post-test check -- surfaces prior fixes when tests fail
• Negative results -- records what you investigated and didn't find, so future sessions don't repeat the search
• Research context -- shows what's been tried, what failed, and active strategies before starting work

# Record a finding
attest_learned("ClaimInput API", "must use provenance= dict, not source_id= kwarg", "warning")

# Record a dead end
attest_negative_result("RustStore", "streaming query API", "read all 27 PyO3 methods")

# Get research context before starting work
attest_research_context("bincode serialization")
# Returns: dead ends, prior findings, active strategies, prior investigators

84 MCP tools total. Works with Claude Code, Cursor, Windsurf, and Codex.

Multi-Agent Collaborative Research

Multiple agents can collaborate on knowledge expansion -- submitting research, claiming tasks from a gap-driven queue, and sharing negative results so failed investigations prune the search tree for everyone.

from attestdb.infrastructure.agents import (
    register_agent, submit_research, submit_negative_result,
    generate_tasks, claim_task, complete_task, get_task_context,
)

# Register agents
register_agent(db, "agent-alpha", capabilities=["literature"], model="gpt-4o")

# Submit research with gaps discovered
submit_research(db, "agent-alpha", claims=[...],
    topic="tp53", gaps_discovered=["mutation spectrum in rare cancers"])

# Record what didn't work (prunes the search tree for other agents)
submit_negative_result(db, "agent-alpha",
    subject=("tp53", "gene"),
    hypothesis=("direct kinase activity", "mechanism"),
    search_strategy="PubMed + STRING + Reactome")

# Task queue auto-generates from knowledge gaps
tasks = generate_tasks(db)  # Entities with 3+ negative results get deprioritized

# Get context before starting (dead ends, strategies, prior findings)
context = get_task_context(db, "tp53")

# Federation: sync claims between instances via NDJSON
export_claims_since(db, since_ns=0, stream=output_file)
import_claims_from_stream(db2, stream=input_file)

Content-addressed claims (content_id = SHA-256 of subject+predicate+object) enable conflict-free dedup across federated instances. Same fact from different agents = automatic corroboration.

Rust Backend

The Rust storage engine powers all operations (1.3M claims/sec ingestion, 8us entity query). Install separately if not bundled:

pip install attest-py

Attest uses the Rust backend automatically when available. LMDB storage: instant crash recovery, concurrent readers, O(1) memory open for any database size.

Core Capabilities

• Provenance-tracked claims -- every fact has a source chain
• Confidence scoring -- Tier 1 (direct evidence) + Tier 2 (corroboration)
• Retraction with cascade -- db.retract("source_123") propagates downstream
• Time travel -- db.at(timestamp) for point-in-time views
• Graph algorithms -- PageRank, betweenness centrality, SVD embeddings, causal prediction
• Contradiction resolution -- evidence-weighted voting across corroboration, source tier, recency
• Snapshot/Restore -- db.snapshot(path) and AttestDB.restore(path)
• MCP server -- 84 tools for AI agent integration
• Embedding index -- HNSW similarity search via usearch
• Audit chain -- tamper-evident Merkle hash chain on append-only log
• Multi-agent research -- agent registration, task queue, federation
• Negative results -- first-class claims that prune the search tree

Intelligence Layer (Enterprise)

LLM-powered features (curation, text extraction, chat ingestion, connectors, insight engine) are available in attestdb-enterprise:

pip install attestdb-enterprise

CLI

attest stats my.db              # Show database statistics
attest query my.db BRCA1        # Query knowledge around an entity
attest schema my.db             # Show knowledge graph schema
attest serve --port 8892        # Start MCP server
attest ingest file.json --db my.db  # Ingest claims from file
attest-mcp install              # Install MCP for your coding tool
attest-mcp metrics              # View hook performance metrics

Common Errors

ProvenanceError -- Every claim needs provenance={"source_type": "...", "source_id": "..."}. This is by design: claims without provenance can't be verified.

ImportError: requires attestdb-enterprise -- Methods like ingest_chat(), ingest_text(), and connect() need the enterprise package. Use db.ingest() for structured claims without it.

ValueError: Unknown vocabulary -- Valid vocab names are bio, devops, ml, ai_tools, or all. Not biology.

database is locked -- Only one process can write to a .attest file. Kill stale processes or use :memory: for testing.

See docs/10_getting_started.md for a full troubleshooting guide.

Documentation

See docs/ for full architecture and design documentation:

• docs/02_architecture.md -- Full technical architecture
• docs/06_api_spec.md -- API contract and validation rules
• docs/07_design_decisions.md -- Critical decisions with rationale
• docs/10_getting_started.md -- Getting started + troubleshooting

Running Tests

pip install attestdb[dev]
pytest tests/unit/ tests/integration/   # ~1100 tests, <75s
cd rust && cargo test                   # Rust unit + golden vectors

License

Business Source License 1.1 -- free to use, modify, and self-host. Converts to Apache 2.0 after 4 years.