Inspiration
AI agents can call tools and make increasingly complex decisions, but their long-term memory is often just retrieved text. A similar past event may be relevant, but that does not mean it is still valid, trustworthy, or safe to act on.
DecisionVault was inspired by one question:
What if agent memory were treated as governed evidence instead of passive context?
Our goal was not simply to help agents remember more, but to help them decide when memory should be trusted.
What it does
DecisionVault is a governed long-term memory layer for AI agents.
It stores verified decision evidence, retrieves semantically relevant experience, and applies governance rules before memory can influence a decision.
In our demo:
Memory OFF → GENERIC_RETRY
Memory ON → REFRESH_PAYMENT_TOKEN
But memory is never blindly trusted. If credible evidence conflicts, DecisionVault safely refuses to act:
CONFLICT_ABSTAIN
action = ABSTAIN
executable = false
It also separates successful execution from verified business success, preventing agents from learning incorrect lessons from incomplete outcomes.
How we built it
We built DecisionVault in Python with:
- CockroachDB Cloud as the authoritative long-term memory store
- Distributed Vector Indexing for semantic retrieval
- Cloud Managed MCP Server for managed database access
- AWS Lambda for the deployed application
- NVIDIA
nv-embedqa-e5-v5for semantic embeddings meta/llama-3.1-8b-instructfor explanation-only output
A key design principle is that semantic similarity and trust are separate. Relevant memories must still pass checks for provenance, verification, revocation, supersession, recency, and conflict before reaching the decision layer.
Challenges we ran into
The hardest problems were not storing or retrieving memory, but governing it safely.
We had to handle cases where:
- stale or revoked memories crowded out valid evidence
- multiple agents produced conflicting conclusions
- newer evidence superseded older memories
- concurrent updates created consistency risks
- an API call succeeded but the real business outcome was still unknown
These cases pushed us toward fail-closed behavior and explicit abstention instead of forcing a decision.
Accomplishments that we're proud of
DecisionVault evolved from a memory prototype into a production-oriented trust layer.
Our final system demonstrates:
- governed semantic memory
- preserved evidence provenance
- conflict-aware abstention
- verified outcome learning
- multi-agent evidence handling
- separation between explanation and execution authority
Final validation includes:
257 / 257 tests PASS
14 / 14 semantic benchmark cases PASS
The result is a system that can both use memory when evidence is strong and refuse memory when evidence is unsafe.
What we learned
We learned that agent memory is not mainly a storage problem. It is a decision integrity problem.
Vector search answers:
What past experience is relevant?
Governance must answer:
Which of that experience is still allowed to influence this decision?
Reliable long-term memory therefore needs provenance, verification, lifecycle control, conflict handling, and safe write-back of outcomes.
What's next for DecisionVault
Next, we want to expand DecisionVault into reusable trust infrastructure for multi-agent systems.
Planned improvements include stronger evidence attribution, richer temporal reasoning, better governance observability, and stronger external outcome verification.
Our long-term goal is simple:
AI agents should not only remember — they should know when their memories are safe to trust.
Log in or sign up for Devpost to join the conversation.