Sign InOpen Brain
AI EngineerVideoSource Linked

No Memory, No Harness: Why the Database Is the Last Line of Defense — Kay Malcolm, Oracle

Oracle’s team used a shared memory broker to carry agent context across people and sessions. The useful pattern is durable, queryable team memory; the database choice is vendor-framed.

AI Engineer · Sep 14, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Oracle’s distributed team found that code crossed time zones without the reasoning that produced it. Its Py memory broker uses the **Oracle Agent Memory SDK** and an autonomous database to share conversations and retained context across team members.

Practical Implication

For team coding agents, persist more than final diffs: retain selected facts, decisions and procedures with retrieval suited to the task. Distinguish **five memory types**—short-term, long-term, episodic, procedural and semantic—rather than treating every transcript as one undifferentiated store.

Agent-Ready Context
Oracle’s distributed team found that code crossed time zones without the reasoning that produced it. Its Py memory broker uses the **Oracle Agent Memory SDK** and an autonomous database to share conversations and retained context across team members.

For team coding agents, persist more than final diffs: retain selected facts, decisions and procedures with retrieval suited to the task. Distinguish **five memory types**—short-term, long-term, episodic, procedural and semantic—rather than treating every transcript as one undifferentiated store.

The talk argues for placing relational, JSON, graph, vector and text data in one **26AI database**, but it presents no comparative evaluation against filesystem memory or separate stores. Retention policy, access control, retrieval quality and deletion behavior remain unspecified.
Connected Context · Feed7 Judgment

This reframes team memory as shared infrastructure and usefully separates five kinds of retained context, but the candidates expose the missing operating policy. A unified database does not by itself decide what deserves retention, suppress harmful combinations, survive model migrations or satisfy access and deletion requirements. The durable contribution is the memory taxonomy and cross-team continuity problem, not evidence that one storage architecture solves them.

MemLens: A Value-Aware Memory Management System with Interactive Analytics for LLM-based AgentsMemLens addresses the unspecified retention policy by scoring individual records and exposing quality, latency and token tradeoffs instead of preserving all context equally.MeClear: Cooperative Game-Theoretic Attribution and Risk-Aware Memory Clearance for Long-Horizon LLM AgentsMeClear shows that successful retrieval is insufficient: combinations of stored memories can harm a task and may require reversible, query-scoped suppression.Does Your Agent's Memory Survive a Model Upgrade? A Controlled Study of Memory PortabilityThe portability study adds a lifecycle requirement absent from the database proposal: unchanged stored memory and mixed embeddings may fail after a model upgrade, so raw histories and migration tests matter.thedotmack/claude-memClaude-Mem provides an inspectable filesystem-oriented alternative for cross-session continuity, making the talk’s untested comparison against separate or local stores directly relevant.
Context Map
agentcodingdata#agent-memory#retrieval#enterprise
Uncertainty
The talk argues for placing relational, JSON, graph, vector and text data in one **26AI database**, but it presents no comparative evaluation against filesystem memory or separate stores. Retention policy, access control, retrieval quality and deletion behavior remain unspecified.