CRDB Agentic Memory

Agent memory on CockroachDB that provably changes actions

Comment

Inspiration

Agents forget. Every restart wipes the operational rules a team has already taught them — "back up the schema before prod deploys" gets re-learned the hard way. We wanted the smallest honest system where a memory demonstrably changes an agent's next action, across process restarts, on infrastructure a solo developer can actually run.

What it does

CRDB Agentic Memory is a persistent, tenant-scoped memory layer on CockroachDB plus a closed-loop demo agent:

  • Store/recall: 384-d embeddings in a memories table; CockroachDB's distributed vector index returns the nearest neighbours for this agent only (prefix-scoped by agent_id), then the application reranks by recency (true half-life) and importance.
  • Memory changes action: a deploy-assistant agent recalls constraints before planning. In session A an operator teaches it two rules; the process exits; in session B — a NEW process and a NEW task — the plan gains a pre-step (schema backup) and post-step (notify #ops), each audited in an actions table with the exact memory ids that shaped it.
  • Self-check via the Managed MCP Server: the agent inspects its own memory infrastructure (query plans, recall audit log) through CockroachDB Cloud's hosted MCP endpoint with a service-account API key — no SQL credentials in the agent.
  • Deployed on AWS: the agent runs as an AWS Lambda container image (embedding model baked in at build time) behind a public Function URL — the demo URL below hits Lambda, which talks to CockroachDB Cloud (aws-ap-southeast-1, same region).

CockroachDB tools used (rule: at least 2)

  1. Distributed Vector Indexing — CREATE VECTOR INDEX on (agent_id, embedding); EXPLAIN on the exact application SQL at 10,000 rows shows vector search ... memories_embedding_idx.
  2. CockroachDB Cloud Managed MCP Server — the agent's self-check calls explain_query / select_query on https://cockroachlabs.cloud/mcp, authenticated by a service-account API key (Cluster Operator role).

AWS services used (rule: at least 1)

AWS Lambda (container image, 2048MB) + Function URL (AuthType NONE — judges can curl it anonymously) serving the agent; the CockroachDB Cloud cluster itself also runs on AWS (ap-southeast-1).

Try it (no auth needed)

Demo URL: https://22ui2md5g3jzumyog776yhmdhy0kvkrh.lambda-url.ap-southeast-1.on.aws/

  • GET returns a canned demo task.
  • POST {"task": "your task here"} plans an arbitrary task against the same memory.

How we built it

Python. memory.py is 100 lines of psycopg2 — no ORM — implementing two-stage retrieval: ANN candidates from the vector index, then a weighted rerank (0.75 cosine similarity + 0.15 recency + 0.05 importance; weights and formula documented exactly as implemented). fastembed (BAAI/bge-small-en-v1.5) for embeddings. The MCP interface uses FastMCP over stdio for local agent demos; the Managed MCP client is stdlib-only (urllib) speaking streamable-HTTP JSON-RPC.

Evaluation (honest numbers)

10,000 memories across 5 agents on a free CockroachDB Cloud Basic cluster, queried from a consumer laptop cross-region:

  • Recall@10 = 1.00 on 100 fact-targeted paraphrase queries (fetch = 32 / 64 / 128 ablation; 32 suffices)
  • End-to-end p50 = 299 ms, p95 = 313 ms — including query embedding on CPU and the public-internet round trip
  • Bulk insert ~241 rows/s; fresh-connection persistence verified

Methodology note: template paraphrases measure retrieval plumbing at scale, not open-domain generalization. English-only model.

Challenges we ran into

  • The optimizer (correctly) full-scans tiny tables; vector search engages at scale — we verified the real query plan at 10k rows instead of trusting a toy EXPLAIN.
  • A naive benchmark gave Recall@10 = 0.2: the queries didn't identify the target fact, making the task unwinnable. Fixing the benchmark (not the system) was the lesson.
  • Cross-process demos mean nothing survives in memory — which is the point, and also what makes them annoying to script.

What's next

Forget/supersede APIs, per-kind decay, provenance and trust levels on memories, and multi-agent auth on the MCP surface.

Pre-existing work disclosure

Built during the submission period. The memory-layer design draws on the authors' prior personal experiments with agent memory (SQLite-based); all code in this repository was written new for this project.

Built With

  • aws-lambda
  • cockroachdb
  • fastembed
  • mcp
  • psycopg2
  • python
Share this project:

Updates