memledger is the open-source trust layer for multi-agent AI — provenance chains, weakest-link confidence, and the Memory Attribution Integrity (MAI) rubric that runs in any eval framework. Open-source by default, no lock-in. Drops onto AWS managed backends when you need them.
From pip install to a governed memory in a few lines.
Multi-agent systems read each other’s beliefs. memledger is what keeps that from quietly failing at scale.
Every memory tracks its derivation; chains span agents and sessions.
Learn more →Effective confidence bounded by min(declared, chain.min). High-conf claims can’t outscore their weakest ancestor.
Near-duplicates flagged at write; CONFLICTS edges visible in the trust graph.
Learn more →Composite 0–100 score with explainable decomposition (confidence, usage, access, reliability).
Learn more →Memory Attribution Integrity runs in RAGAS today. DeepEval, Phoenix Evals, LangSmith, OpenAI Evals on the roadmap.
Learn more →Right-to-be-forgotten propagates through derivatives — compliance-grade audit trail.
Learn more →memledger sits between your agents and storage, applying provenance, confidence, conflict, and audit guarantees on every operation. Eval frameworks score MAI from above; backends serve from below.
The Console UI surfaces every memory’s chain, conflicts, and Memory Quality Score breakdown — live, on the same memory the agent just used.
One agent writes a low-confidence guess. Another reads it and treats it as ground truth. The bad belief propagates silently.
memledger gates retrieval on chain-bounded effective confidence. Below threshold, the memory is filtered or flagged before your agent ever sees it.
hits = await ml.search(
query="connection pool fix",
namespace="/ops/payment-svc",
top_k=5,
confidence_policy={"min_threshold": 0.5},
)A user invokes GDPR Art. 17. The memory you wrote is fine to delete — but ten derivatives across three agents reference it.
memledger plans the cascade across derivatives, shows you exactly what will be deleted, and executes with a signed audit record.
plan = await ml.plan_cascade(memory_id=incident_id)
result = await ml.execute_cascade(
plan_id=plan.id,
reason="GDPR Art. 17",
requester_id="dpo@org",
)You ship an agent. It works in eval. Three weeks later, retrieval drift has degraded answers and you have no signal.
memledger’s MAI rubric runs in your existing eval stack. Score Memory Attribution Integrity on every release; alert when it slips.
from evaluators.attribution_integrity_ragas import evaluate_mai_ragas
report = await evaluate_mai_ragas(
ledger=ml,
dataset="calibration-30",
)
print(report.summary)Diagnose how agent recommendations went wrong — chain, conflicts, and confidence in one trace.
Gate retrieval on chain-bounded effective confidence. Sane defaults, tunable per workload, no lock-in.
Audit every memory’s provenance and run RTBF cascades through derivatives. Apache 2.0; deploy in your VPC.
Drop memledger under LangGraph, OpenAI Agents SDK, or kagent. The trust layer is framework-agnostic.
A reference implementation of provenance, weakest-link confidence, and MAI you can read end to end.
If two agents share memory, you need attribution. memledger gives you that without rewriting your stack.