Show HN: Detailed Guide to Agent Memory from a 3-Year Cognee Builder

A developer building a local memory layer for AI agents describes a rule set for deciding which temporary context should become durable memory. The rules include verified decisions, stable user preferences, reusable procedures, facts with sources, and corrections that supersede older facts. The post highlights the challenge of preventing stale or intermediate outputs from polluting long-term memory.

A developer reports that traditional agent memory approaches—stuffing context windows or searching history—fail when knowledge graphs become complex, leading to confident but wrong answers. They built a statement graph with typed entities, relationships, and metadata about statements to track truth over time.

Social·Community

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

Kay Malcolm, who runs outbound database product management at Oracle, gave a talk arguing that AI made her distributed team faster individually without improving team productivity because Git captures code changes but not reasoning. She frames the model as a brain, the harness as a body, and memory as the nervous system, then argues memory should live in a single database rather than four specialized stores that force agents to guess at reconciliation. The talk matters to practitioners building agent memory layers, proposing one store covering all five memory types instead of a multi-database stack.

Kay Malcolm, who runs outbound database product management at Oracle, gave a talk arguing that AI made her distributed team faster individually without improving team productivity because Git captures code changes but not reasoning. She frames the model as a brain, the harness as a body, and memory as the nervous system, then argues memory should live in a single database rather than four specialized stores that force agents to guess at reconciliation. The talk matters to practitioners building agent memory layers, proposing one store covering all five memory types instead of a multi-database stack.

A developer running Claude Code and Codex on the same project found semantic memory retrieval worked correctly but agents still acted on outdated decisions — Codex rebuilt a sqlite cache layer after an August switch to duckdb because it retrieved the older, more confidently written decision. A custom recency-scoring resolver (200 lines) failed too, since decision recency differs from write recency. The working fix is one shared memory store (via Vilix AI) across Claude Code, Codex, and Cursor with the rule that the newest explicit write wins.

A developer reports production bugs found after roughly two months of piloting an agent memory system that unifies six sources into one customer context layer. None of the five bugs raised errors, including recency queries returning stale July facts because semantic-plus-keyword search has no date term.