Konshus
AI Memory News

AI agents can keep memories. Can they keep them straight?

Long-term memory has moved from research prototypes into AI products. The harder question is whether agents can update, question and safely forget what they store.

2026-11-07 · The Konshus Team · Sources checked before publishing

Share

Want more news like this?

  • One short email a week with the biggest changes in AI memory — ChatGPT, Claude, Gemini, new companies and research.
  • Every story is checked against its sources before it goes out.
  • No spam: just the weekly news, no sales emails, and we never sell or share your address.

Free. One email a week, nothing else — no Konshus marketing emails, and we never sell or share your address. Unsubscribe anytime in one click.

An AI agent can remember a preference from last month without keeping last month’s entire conversation in front of it. That is the promise of long-term memory: store something now, find it later, and use it when it matters. It sounds simple until the preference changes, the stored account proves mistaken, or someone else manages to plant a convincing instruction in the record.

The research increasingly treats memory as a managed system rather than a bigger prompt. A survey describes three core operations: writing information, managing what has been stored, and reading relevant information back. Another distinguishes memories of particular events from stored facts and procedures—different material that may call for different treatment. Agent-memory survey · Survey of memory mechanisms

That shift matters because persistent memory is already a product feature, while research benchmarks still find substantial weaknesses in long-running conversations. The question is no longer simply whether an agent has memory. It is what the agent chooses to keep, how it resolves change, and whether people can trust or remove what it remembers.

What changed: memory became a system to manage

Early work established two influential approaches. In Generative Agents, simulated characters recorded experiences, retrieved them according to relevance, recency and importance, and formed higher-level reflections to guide later behavior. MemGPT approached the limited context window as a storage-management problem: move information between the model’s working context and other storage rather than expect everything to fit in one prompt. Neither paper established dependable, general-purpose memory, but both helped make the design problem concrete. Generative Agents · MemGPT

There is no single agreed architecture. A recent survey groups approaches including compressed context, retrieval from external stores, reflection, hierarchical management of context and learned memory management. A system might preserve conversation excerpts, make summaries, extract facts or organize records into a structure. Those choices affect what evidence remains available when an agent needs to answer a later question. Survey of autonomous-agent memory

Consider a user who says in January that they prefer morning meetings, then says in June that mornings no longer work. Retrieving either statement is not enough. A useful agent must work out which statement applies, whether the change is temporary, and what to do if it cannot tell. This is why memory research has moved beyond measuring whether a system can find a sentence: the management and use of that sentence matter too.

What the tests actually tell us

The benchmarks do not deliver one definitive memory score. LongMemEval asks 500 questions across information extraction, reasoning over multiple sessions, temporal reasoning, knowledge updates and abstention—knowing when not to answer. Its authors report an accuracy drop of about 30% for the commercial assistants and long-context models they evaluated when information was distributed across sustained interactions. That is a finding from a particular test, not a measurement of every deployed assistant. LongMemEval

LoCoMo tests different demands using long conversations averaging roughly 600 turns and 16,000 tokens, spanning up to 32 sessions. Its authors report that long-context and retrieval-based approaches help but remain substantially behind human performance on their tasks, including long-range temporal and causal understanding. These results help explain why putting more text within reach is not the same as using it correctly. LoCoMo

MemoryAgentBench adds incremental, multi-turn interaction and evaluates capabilities including retrieval, learning during a test, long-range understanding and selective forgetting. Its premise is that a static question about an existing record misses what happens when an agent accumulates information over time. Together, these benchmarks probe different failure modes; their scores should not be treated as interchangeable league-table results. MemoryAgentBench · LongMemEval · LoCoMo

The distinction for readers is between finding evidence and forming a warranted answer. A retrieved passage may be relevant but outdated. Two passages may conflict. Sometimes the right answer is that the record does not say. LongMemEval tests updates, timing and abstention separately for a reason: success at recall alone cannot settle those questions. LongMemEval

Who is building on the idea?

Research groups and system developers are trying to improve the whole process, not just retrieval. The Mem0 paper describes extracting information from conversations, consolidating it and retrieving it later. Its authors report better results than their selected memory and retrieval baselines on LoCoMo, alongside lower latency and token costs than supplying the full conversation. That is a useful comparison within their evaluation, not proof that Mem0 wins across deployments. Mem0

A 2026 paper on MOSAIC likewise reports substantial gains over its baselines on evaluated LoCoMo and HaluMem settings, including conflict handling. The takeaway is that organization and reconciliation can improve measured results. The size of any advantage remains dependent on the test and comparison set; developer-reported results should not be mistaken for an independent verdict on the field. MOSAIC paper

The major product examples show another side of the story: giving people access to persistent memory. OpenAI distinguishes ChatGPT’s saved memories from information referenced from past chats. Its documentation says saved memories are separate from chat history, so deleting a chat does not automatically remove a saved memory, and it describes controls for both forms of reference. That separation is useful to understand—and important to know before assuming that deleting a conversation clears every trace it created. OpenAI Memory FAQ

Anthropic documents Claude memory with controls for viewing and editing it and describes project-scoped memory. Separately, its developer documentation describes a memory tool through which an agent can create, read, update and delete files that persist between sessions. These are meaningful ways to make memory available and manageable, although product documentation cannot establish that an agent will recall the right detail or interpret it correctly every time. Anthropic on Claude memory · Claude memory tool documentation

What remains unsettled

Forgetting is not automatically safer or smarter. Keeping too much can preserve stale or sensitive information; keeping too little can discard evidence needed later. In one 2026 study of memory filtering, removing relevant information was particularly damaging in the tested settings. The study also found that retaining evidence did not ensure a correct answer when incorrect labels steered the model. It illustrates a trade-off, not a universal instruction to store everything. When Should LLM Agents Forget?

A once-accurate memory can become dangerous. A study of spatial memory in vision-language agents reports that, in its tested environment, agents relying on stale memory performed worse on safety than agents given no memory. That narrow result does not predict how a chat assistant will behave. It does show why an agent acting in a changing setting may need to check what it remembers against what it observes now. When Memory Lies

A memory can also be tampered with. A 2026 security survey examines risks across writing, storage, retrieval, execution, sharing and forgetting. It finds defenses particularly underdeveloped around storage, sharing and verified removal, and says none of the benchmarks it reviewed covers the full memory lifecycle. A separate poisoning study reports that more aggressive writing and retrieval made some tested agents more exploitable. Neither result means every memory feature is compromised; both argue against treating a memory write as harmless note-taking. Memory-security survey · Memory-poisoning study

What it means

A practical test of agent memory should ask more than, “Did it remember?” Ask whether it can show where a remembered claim came from, recognize when a newer statement changes it, decline to rely on stale information, and remove or contain a record when appropriate. The security survey identifies provenance, scoped retrieval, rollback and verified forgetting as needs; current benchmarks cover important pieces of performance but not the entire lifecycle. Memory-security survey · MemoryAgentBench

Long-term memory is real as a system capability, and the research gives good reasons to build it carefully. It has not established human-like, dependable remembering. For now, the most revealing question is not how much an agent can store, but how it decides which memories deserve to guide its next action.

Our take (opinion)

The industry should stop treating persistence as the finish line. OpenAI and Anthropic have done something useful by documenting distinct kinds of memory and giving users ways to manage them. Researchers, meanwhile, have made recall failures and update problems much easier to test. Both are progress.

Our concern is that a correct-looking memory can gain authority simply by surviving. An agent may use an old preference without checking its date, or treat a stored claim from an untrusted source as if the user said it. More storage and better search will not resolve those questions on their own.

The better standard is accountable memory: an agent should be able to distinguish a quote from an inference, preserve enough context to handle corrections, and make deletion meaningful. Benchmarks should test those abilities together, over time. Until then, claims that an agent “remembers you” deserve a follow-up: what happens when its memory is wrong?

Questions

Is long-term AI memory the same as a larger context window?

No. A larger context window lets a model consider more material at once. Long-term memory generally involves storing information outside the current interaction, managing it, and retrieving selected items later.

Do benchmarks show that AI agents remember reliably?

No. Benchmarks show that memory designs can help, but they also identify weaknesses in multi-session reasoning, updates, timing and knowing when to abstain. Different tests measure different abilities.

If I delete a ChatGPT conversation, does that delete its saved memory?

Not automatically. OpenAI says saved memories are stored separately from chat history and describes separate controls for managing them.

Why can keeping more memories make an agent less safe?

Stored information can become stale, sensitive or untrustworthy. Research has found risks from stale spatial memories and from adversarial information entering agent memory; those findings depend on the systems and settings tested.

Sources

  1. A Survey on the Memory Mechanism of Large Language Model Based Agents
  2. A Survey on Memory Mechanisms in the Era of LLMs
  3. Generative Agents: Interactive Simulacra of Human Behavior
  4. MemGPT: Towards LLMs as Operating Systems
  5. Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Open Problems
  6. LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory
  7. Evaluating Very Long-Term Conversational Memory of LLM Agents
  8. Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions
  9. Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory
  10. Accurate and Efficient Long-Term Memory for LLM Agents
  11. OpenAI Memory FAQ
  12. Bringing memory to teams
  13. Claude memory tool documentation
  14. When Should LLM Agents Forget?
  15. When Memory Lies: An Empirical Study of Spatial Memory Staleness in VLM Agents
  16. A Survey on Long-Term Memory Security in LLM Agents
  17. From Untrusted Input to Trusted Memory

The AI memory news, once a week.

If you're interested in AI memory, want your own AI to remember you better, or just want to stay up to date, this is the newsletter for you. Every week we comb through the news for the big shifts in AI memory — new companies, new features from ChatGPT, Claude and Gemini, new research and new ideas — and sum up what matters in one short email.

For people who use AI every day, builders and founders, and anyone curious where AI memory is going.

Free. One email a week, nothing else — no Konshus marketing emails, and we never sell or share your address. Unsubscribe anytime in one click.

About Konshus: a private memory Vault that works across ChatGPT, Claude, Gemini and more. Learn more.