Agent Memory Requirements

Type: kb/types/note.md · Status: current · Tags: agent-memory, context-engineering

This directory expands the requirements inventory derived in Designing a Memory System for LLM-Based Agents. Each note describes one requirement in a form an agent can load independently while designing, reviewing, or comparing memory systems.

The requirements are organized by operation, not by implementation layer. A system may satisfy them through files, databases, runtime services, prompts, skills, review workflows, or learned policies. The point is to make the needed capability and failure mode explicit.

Requirements

  1. Create memory directly - write useful memory artifacts while understanding, provenance, and caveats are still live.
  2. Import external knowledge into internal form - transform existing project knowledge into typed, linked, trusted system artifacts.
  3. Preserve evidence without loading history - keep traces for extraction and audit without making history the default context.
  4. Use trace-derived extraction as meta-learning - mine traces after the fact while respecting signal quality and review.
  5. Serve multiple consumers - expose different surfaces for acting, scheduling, review, learning, governance, and active work.
  6. Activate behavior-changing memory - fire relevant lessons before repeated mistakes or high-risk actions.
  7. Promote only when value exceeds cost - move candidates into durable artifacts only when future value justifies maintenance.
  8. Keep compiled views aligned - prevent cues, prompt files, indexes, and generated surfaces from drifting into independent authority.
  9. Retire, redact, supersede, and relax memory - manage forgetting, revision, sensitivity, staleness, and temporal validity.
  10. Make authority explicit - define who can read, write, promote, activate, enforce, revise, and retire memory.
  11. Evaluate memory by effects - test whether memory changes future tasks, artifacts, answers, and behavior.

Use

Use the synthesis note when you need the derivation. Use these notes when you need to inspect, implement, or score one capability.