Three separate things landed in the past week that point the same direction: agent memory is becoming an explicit, inspectable store rather than a summarization step buried in the loop.
The most concrete is OpenAI's GPT-6 Astra. Per the llm-releases tracker, the model ships with an updated Codex harness that keeps searchable notes across context windows instead of lossy compaction, alongside a 1M-token context window. It's available as `gpt-6-astra` in the OpenAI API and on Amazon Bedrock, with a GPT-6 Astra Pro tier for ChatGPT Pro, Business, and Enterprise that's off by default at launch and enabled per workspace. LLM Stats lists it at $10 input / $50 output per 1M tokens, with cached input at $1 and cache writes at $12.50.
The harness change matters more than the context number. Compaction is a summarizer you don't control, running on a schedule you didn't choose, silently dropping the constraint the user gave you forty turns ago. Notes you can search are a store: they have a schema, they can be diffed, they can be audited when the agent gets something wrong. If you've been building a memory layer whose main job was cramming the right things into the prompt before the window filled, that job just moved. What's left — and it's the harder part — is deciding what deserves to persist, how it gets invalidated, and who can read it.
Research the same week says the retrieval side isn't a model-scale problem. A paper published August 30 by Ming Wu and Pengyuan Zhu introduces Agent Zero Memory, reporting 95.60% on LongMemEval and 93.60% on LoCoMo, with a companion cost analysis finding near-state-of-the-art quality at up to 20x lower cost per query. The design premise: not one general-purpose store but three parallel, purpose-built structures, with every fact carrying proof of where it came from. Provenance as a first-class field, not a metadata afterthought.
Mem0's State of AI Agent Memory 2026, out two days ago, frames the same shift from the ecosystem side — memory systems moving beyond pure vector similarity, with vector retrieval finding semantically similar facts and graph-style retrieval traversing entities and relationships, and neither sufficient alone. The report also points at Neptune, Cassandra, and Valkey backends and local embedding via FastEmbed for teams that can't ship data off-box.
Two cautions. First, longer windows don't shrink the problem — Google's Gemini 3.8 Flash (Sep 2) came with roughly 30% more average output tokens per task, near 48k, which raises cost per task even at unchanged token pricing. Agentic loops are getting longer, not tighter. Second, an externalized note store is a new staleness surface: facts that used to die with the context now survive indefinitely unless you build eviction.
If you own a memory layer, the useful work this quarter is provenance, invalidation, and a domain eval set — not a bigger backbone.