In the first half of 2026 the agent-memory field shifted from "memory = conversational recall" to "memory = self-evolving knowledge store" — simultaneously in academia, open source, and products

agent-memorypeer-fieldbenchmarksanthropic-dreamsperplexity-brain2026-shift

The shift is visible in three lanes at once:

• Academia (arXiv 2026): a field-defining survey, "From Storage to Experience" (2605.06716), stages agent memory as Storage → Reflection → Experience. New benchmarks test compounding, not recall: MemoryArena (2602.16313 — memory acquired then relied on across interdependent sessions), EvoMemBench (2605.18421 — 15 methods, cross-episode). Closest to a verified store: U-Mem (2602.22406), where the agent actively acquires and validates knowledge with tiered validation.
• Open source: Karpathy's LLM Wiki gist (claim-karpathy-llm-wiki-gist-canonized-compounding-pattern) and Nakajima's ActiveGraph (claim-activegraph-rebuilds-babyagi-on-event-log) — the two people whose January-2026 priors most needed updating overturned them themselves, in the same spring.
• Products: Anthropic shipped platform-level memory stores (public beta, header managed-agents-2026-04-01) plus Dreams (header dreaming-2026-04-21): "A dream reads an existing memory store alongside past session transcripts, then produces a new, reorganized memory store: duplicates merged, stale or contradicted entries replaced with the latest value, and new insights surfaced" (docs, direct fetch; the closing clause "with the latest value, and new insights surfaced" was dropped at promotion without an ellipsis — restored 2026-09-11 audit. Post-capture drift, same re-fetch: the page now labels Dreaming a "research preview", and memory-store endpoints have moved to an agent-memory-2026-07-22 header; sessions still use managed-agents-2026-04-01). Perplexity's Brain (2026-06-18, research preview) is described as "an LLM wiki that is automatically loaded onto the agent sandbox" [unverified verbatim — triangulated via T3] (2026-09-11 audit: perplexity.ai/hub/blog 403s WebFetch on both Brain posts; a search snippet of Perplexity's own post reads "an LLM wiki that's automatically loaded onto the agent sandbox" — still no direct read, flag stands).

What did not shift: the consumer deep-research buttons (ChatGPT, Gemini, Claude Research) still neither read nor write a compounding research-findings store; the STORM-class survey pipelines and the Sakana AI-Scientist line remain one-shot with no revisit protocol (priors confirmed against repos/papers 2026-07-06). The distinction that survives: chat/preference memory is now universal; a compounding knowledge store is still the exception; a compounding verified store is nearly nonexistent.