GroundTrace Memory

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

Inspiration

Agentic memory creates a deceptively difficult trust problem: retrieval is not the same thing as authority.

An incident-response agent may retrieve a highly similar past incident, but that memory could be unverified, stale, based on a failed action, or missing provenance. GroundTrace Memory was built around one principle:

Retrieved memory is a candidate, not authority.

The goal was to make memory useful without allowing semantic similarity alone to determine consequential action.

What it does

GroundTrace Memory stores resolved incidents, actions, outcomes, verification state, provenance and embeddings in CockroachDB.

When a new incident arrives:

Amazon Bedrock Titan Text Embeddings V2 creates a 1,024-dimensional embedding.
CockroachDB Distributed Vector Indexing retrieves semantically similar memories within the same tenant.
A deterministic admission gate evaluates every retrieved candidate as RELY, HOLD, or REJECT.
Only a RELY memory may supply a bounded recommended action.
The complete decision trace is persisted in CockroachDB for later inspection.

The gate considers verification, prior outcome, freshness, similarity and provenance. A similar memory can therefore be retrieved successfully while still being prevented from influencing the agent.

How we built it

The submission is a deliberately narrow end-to-end vertical slice.

AWS Lambda exposes the memory operations and uses Amazon Bedrock Titan Text Embeddings V2 for semantic embeddings. CockroachDB Cloud acts both as the transactional system of record and the semantic memory layer.

Resolved incidents are stored transactionally alongside their embeddings and evidence. Recall is tenant-scoped and uses CockroachDB Distributed Vector Indexing to find candidate memories.

The admission layer then applies deterministic rules:

unverified memory → HOLD
unsuccessful prior outcome → REJECT
memory older than 180 days → HOLD
cosine similarity below 0.70 → HOLD
missing provenance → REJECT
otherwise → RELY

Every recall persists a decision trace containing candidate IDs, similarity scores, admissions, reasons, recommendation and overall status.

CockroachDB Cloud Managed MCP Server is also used as a read-only operator/judge inspection path for the stored memories and decision traces.

Challenges we ran into

The central design challenge was separating relevance from reliability. Vector search is excellent at finding semantically related experience, but similarity alone does not establish that an agent should act on it.

The build also had to preserve evidence across the full path: embedding, retrieval, admission, recommendation and persisted decision trace.

During final deployment verification, a clean AWS SAM deployment exposed a missing packaged root CA certificate for the CockroachDB TLS connection. That was diagnosed from Lambda logs, fixed reproducibly in the repository, redeployed and revalidated end to end.

Accomplishments that we're proud of

The finished system demonstrates a complete live path:

remember → embed → persist → vector recall → RELY/HOLD/REJECT → bounded recommendation → persisted trace

The final live verification confirmed authenticated API access, CockroachDB connectivity, successful memory storage, semantic recall and a RELY admission producing the expected recommendation.

The same stored incident memory and decision trace were then inspected independently through the CockroachDB Cloud Managed MCP Server using read-only tools.

What we learned

The most important lesson is that agentic memory benefits from an explicit admission boundary.

Similarity answers: What looks relevant?

The admission gate answers: What is this agent actually allowed to rely on?

Combining semantic retrieval with transactional evidence and deterministic admission makes the resulting memory system much easier to inspect, constrain and reason about.

What's next

The same pattern can be extended with configurable admission policies, richer temporal reasoning, source-specific trust rules and additional evidence types while keeping the core invariant intact:

memory may inform action only when its warrant survives inspection.