Evolving summaries beat append-only memory — rewrite profiles, don't accumulate facts

The critical innovation in file-based agent memory is a function that rewrites summaries rather than appending to them. When a user switches jobs, the system updates their profile — it doesn’t add a second employer alongside the first. This handles contradictions naturally: the summary always reflects current truth, while the raw resources layer preserves the historical record for audit.

The write path makes this concrete through four stages: resource ingestion, extraction, batching, and evolving summaries. This is the mechanism behind what Persistent agent memory preserves institutional knowledge that walks out the door with employees describes at scale — the “daily changelogs and decision logs” only stay useful if they’re periodically consolidated into evolving summaries. It’s also why Spec files are external memory that survives context resets works: spec files are living documents that get rewritten as understanding deepens, not append-only logs that grow stale. Glean’s trace learning extends this pattern by recognizing that Two-tier agent memory separates organizational workflow knowledge from individual user preferences — deployment-level summaries capture shared workflow strategies while user-level summaries capture individual preferences, each evolving independently. Yet evolving summaries are only one piece of the puzzle: Knowledge evolution is the biggest unsolved problem across all graph architectures reveals that across 10+ knowledge graph architectures, nobody has fully solved graph maintenance — summaries handle profile-level evolution, but atomic insights still lack pruning, merging, and contradiction detection. Hermes offers a complementary resolution to the rewrite-vs-preserve tension in Compression should be a forking lifecycle event, not a destructive rewrite — rewrite the working summary for the live context, but fork a child session so the pre-summary lineage stays recoverable rather than overwritten.