AttestDB

An embedded database for claims with sources, confidence scores and timestamps. The Python API uses a Rust storage engine backed by LMDB.

This is the 0.2.2 release candidate. Brain features below require this candidate; they are not a claim about older published packages.

Start with AutoCompany

Give your existing Dot, Claude Code, Codex or other capable client this prompt:

Help me set up my company on Fabric. Read the AutoCompany setup skill at skills/autocompany-setup/SKILL.md in this checkout. Ask what my company does, which tools we use and which workflow to automate first. Help me connect it, create the first useful work item, and show what is ready and what needs attention.

The setup guide explains the portable skill and private workspace creator. This is a guided setup kit; live services and account connections are verified individually rather than implied by installation.

Choose a Fabric reference implementation:

Priority	Fabric reference implementation	Purpose	Readiness
1	Agent Memory Fabric (Brain)	Persistent findings, failed approaches and session outcomes	Tested release candidate
2	Software Engineering Fabric (Attest Swarm)	One synthetic builder, real tests, Brain memory and replay	Runnable offline example
3	Compliance Fabric	Find uncovered obligations; reopen gaps after evidence withdrawal	Runnable synthetic example
4	Sales and Marketing Fabric (AutoFunnel)	Sourced account briefs, drafts and approval-gated local staging	Runnable offline example
Later	Research Fabric	Investigate unanswered questions and conflicting sources	Proposed
Later	Support and Incident Fabric	Track diagnostic evidence and verified resolutions	Proposed
Later	Supplier Assurance Fabric	Review missing or stale supplier evidence	Proposed

A Fabric names the application domain; Brain, Attest Swarm and AutoFunnel name the corresponding implementations.

See the architecture and delivery plan for boundaries, acceptance gates and what each label means. The planned references are not included applications or supported hosted services.

pip install attestdb

from attestdb import AttestDB

db = AttestDB("example.attest", embedding_dim=None)
try:
    db.ingest(
        subject=("service-a", "service"),
        predicate=("depends_on", "relates_to"),
        object=("cache", "service"),
        provenance={"source_id": "deployment-1", "source_type": "manual"},
        confidence=0.9,
    )
    assert len(db.claims_for("service-a")) == 1

    result = db.retract("deployment-1", reason="Superseded deployment record")
    assert result.retracted_count == 1
finally:
    db.close()

Agent Memory Fabric

Brain is the local agent-memory implementation.

Install the MCP extra, then configure your coding client:

pip install 'attestdb[mcp]'
attest brain install --tool claude
# Or: attest brain install --tool codex

Restart the client. Ask it to call attest_learned("auth.py", "Check session expiry before accepting a token", "tip"), then attest_check_file("auth.py"). Findings remain after the MCP server restarts. attest_negative_result, attest_get_prior_approaches and attest_session_end store failed approaches, recall topics and record reported outcomes.

Storage is local and requires no model API key. Topic recall is lexical; file recall is exact. Records are agent-reported evidence, not independently verified truth. The public installer adds MCP configuration without lifecycle hooks or semantic recall, and preserves other client settings.

Software Engineering Fabric

Attest Swarm is represented here by an offline synthetic example.

From a source checkout with the candidate AttestDB package installed:

python examples/swarm_reference.py --run-dir /tmp/attest-swarm-demo --db /tmp/swarm-memory.attest
python examples/swarm_reference.py --run-dir /tmp/attest-swarm-demo --replay

A synthetic builder repairs a tiny fixture; real subprocess tests check the result, then public Brain memory stores the evidence. One builder attempt is allowed per run directory. Failed checks or an explicit review policy require human review. This example makes no model calls. See the walkthrough and limits.

Sales and Marketing Fabric

AutoFunnel is represented here by an offline synthetic example.

From a source checkout with the candidate AttestDB package installed:

python examples/autofunnel_reference.py

Synthetic account and product facts become a cited brief and a template draft. Only an unchanged, approved draft with explicit demo consent and no suppression can enter the local outbox. Withdrawing required evidence blocks further staging. No messages are sent and no model calls are made. See the walkthrough and boundaries.

Compliance Fabric

The current implementation is a synthetic evidence-coverage example.

From a source checkout with AttestDB installed:

python examples/compliance_evidence.py

The example starts with two uncovered, invented obligations. Adding synthetic evidence covers one; withdrawing its source reopens that gap. It demonstrates evidence-link coverage, not certification, legal compliance or evidence acceptance.

The open source package contains the claim engine and selected basic connectors. Proprietary models, model training and evaluation assets, enterprise connectors, plugins and hosted services are outside this release.

For a source checkout, run the Python tests with pytest tests/ and the storage engine tests with cargo test --manifest-path rust/Cargo.toml --all.

Python 3.11 and 3.12 are the release test targets. If a matching native wheel is unavailable, building attest-py from source requires a Rust toolchain.

CI and release validation run on private runners. See release validation.

Licensed under Apache-2.0.