{
  "id": 19633393,
  "title": "What Agent Memory Should Keep, What It Should Drop — and Why That Is Harder Than It Sounds",
  "url": "https://urgent.news/2026/11/12/what-agent-memory-should-keep-what-it-should-drop-and-why-that-is",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-12T10:02:22.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/zxpmail/what-agent-memory-should-keep-what-it-should-drop-and-why-that-is-harder-than-it-sounds-10b"
  },
  "original_language": "en",
  "account": "Trust, not capacity, is the main tension in agent memory systems. Garbage is not merely useless information; it is information that gets treated as knowledge when reused, even if it is false or stale. The challenge lies not in taxonomy, but in memory's tendency to rot. This is a design note, not a benchmark paper.\n\nMetadata, rather than DB column comments, is the long-lived rule layer that the system trusts by default, including safety constraints, experience rules, environment defaults, and representation structure. The safety constraints and experience rules are expanded further, with safety constraints being unconditional and safety-relevant, while experience rules need conditions to avoid over-generalization.\n\nThe process of correcting errors is simple: return to the raw execution record, demote, rewrite conditions, or mark the entry as overturned. All memory layers move together: session → experience → metadata, with demotion, expiry, and a path back to raw logs when the environment changes.\n\nUnconditional rules, like safety constraints, are reviewed against whether the design still holds, while experience rules are reviewed against whether the conditions still cover the field. Four criteria govern the order of prioritization: verifiability, clear conditions, reusability, and retrievability.\n\nDerived facts should be evaluated based on their derivation cost, not banned outright. Cheap and stable derivation should be computed each time, while expensive or external-state-dependent derivation may be stored as a cache, validated on read. Rules are generators, and caches are snapshots. A snapshot without an invalidation path is future garbage.\n\nThe three layers—session, experience, and metadata—move together, with the session holding evidence, the experience layer holding current-task items, and metadata holding reviewed and approved entries. Unverified entries that hold with clear conditions become experience rules, while falsified entries are kept as counterexamples. Safety-constraint metadata moves only with design changes and human review, never through auto-promotion.\n\nDemotion and expiry are routine processes, not incident responses. The frequency of review depends on risk assessment, with multiple verifications required for high-risk work. Metadata entries must include content, type, basis, and review conditions, and must specify who approves. Periodic review should be based on failure cost, with low-cost experiences reviewed weekly or per sprint, and high-risk metadata reviewed only when a design change occurs.",
  "summary": "What Agent Memory Should Keep, What It Should Drop — and Why That Is Harder Than It Sounds This piece turns an earlier set of principles into an operable mechanism draft, with evidence bounds marked. (Hard flaws in the old draft are listed at the end.) Audience: people building agent memory systems. Basic concepts are not taught here; conclusions carry premises and scope. Non-engineering readers…",
  "key_points": [
    "Trust, not capacity, defines agent memory tension.",
    "Metadata layers include safety constraints and experience rules.",
    "Demotion and expiry are routine, not incident responses."
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}