Signal SentrySearchUpdated Nov 20, 05:30 UTCAM Drop
Frontier watch · Orchestration & memory

Agent memory across sessions and resets

How do agents keep state and memory across sessions, context resets and handoffs, and which memory designs hold up over days of work?

Why it matters

Work that runs for days outlives any context window. Whether an agent can pick up where it left off, and what it is allowed to remember, decides if long jobs finish and creates a new store of data to secure and govern.

State of play

Newest first. Each claim is our summary of the linked primary source, with a short quote from it.

  1. ResearchOct 14, 2026Source

    A survey preprint on multimodal agent memory sorts memory management into three tiers: static memory, dynamic memory updated by a stable policy, and self-evolving memory whose update policy itself changes.

    Axis II, Memory Management, distinguishes three tiers: T1 Static Memory reuses fixed memory; T2 Dynamic Memory updates memory through a stable policy; and T3 Self-Evolving Memory updates memory through an evolving policy.

    What it does not show: Survey and taxonomy; no new experimental results.

  2. ResearchOct 10, 2026Source

    A preprint on LLM-run robot teams replaces the orchestrator's growing interaction history with a state summary rebuilt from a semantic map at every call, one of three fixes for missions that never end.

    A Markovian state summary, rebuilt from the semantic map at every invocation, replaces the growing interaction history.

    What it does not show: Simulation only; the reported 98% vs 89% civilians rescued and six times fewer tokens are for the whole framework, not the summary alone.

  3. ResearchSep 17, 2026Source

    A preprint describes an agent harness that keeps a bounded summary file per time-scale level; over a ten-day campaign it reproduced a published RL result and kept the thread across every context reset.

    We report on a ten-day campaign in which an agent built on this architecture reproduced a published reinforcement-learning result with a human attending once a day, and show (1) the agent kept the thread across every context reset and session boundary of the campaign

    What it does not show: Preprint; one campaign reported by its authors, with a human attending once a day.

  4. DocumentedSep 10, 2026Source

    OpenAI's API changelog says its Agents API runs a managed Codex harness in which OpenAI handles session orchestration, context compaction and recovery for durable agent sessions.

    Build agents with a managed Codex harness while OpenAI handles session orchestration, context compaction, and recovery.

    What it does not show: Public beta per the same entry; it describes the feature, not how well compaction preserves state.

  5. AnnouncedMay 19, 2026Source

    Anthropic says dreaming, a research preview in Claude Managed Agents, is a scheduled process that reviews past agent sessions and memory stores and curates memories so agents improve over time.

    Dreaming is a scheduled process in Claude Managed Agents that reviews agent sessions and memory stores, extracts patterns, and curates memories so agents improve over time.

    What it does not show: Vendor post; research preview with access by request; the customer result it cites (~6x completion rates) is the customer's own test, method not shown.

What would change the picture

Signals we are watching for. Our own reading of where the thread could go: not predictions, and not sourced claims.

  • A major vendor makes cross-session agent memory, or automatic memory curation like dreaming, generally available instead of beta or preview Would suggest: Persistent agent memory moving from experiments into production infrastructure
  • A shared benchmark compares memory designs on multi-day agent runs that must survive context resets and handoffs Would suggest: Agent memory becoming measurable rather than judged by single anecdotes
  • One vendor's agent can read and write a memory store created by another vendor's harness through a published format Would suggest: Agent memory becoming portable rather than locked to one platform
  • A vendor or researcher documents a production case of poisoned or stale agent memory driving a wrong action Would suggest: Memory turning into a security and governance surface that needs controls

From the Drops

AgentsResearch

Preprint: a weaker agent did better with 29 of its learned guidelines than all 199

Researchers whose memory pipeline runs in IBM's CUGA agent report gpt-oss-120b went from 39.9% to 51.2% on AppWorld with all 199 guidelines, and 55.4% with about 29. A strong model kept gaining as guidelines grew.

Why it matters: More memory is not free accuracy. Size agent memory per model, and keep provenance when lessons are merged; the authors say theirs was silently dropped.

From Drop #097 · Fri, Nov 20, 2026 · Morning
AgentsResearch

Preprint: HiVe-Mem cuts multimodal agent memory tokens by about 90%

Researchers report HiVe-Mem links memory episodes in a graph and picks the evidence form per query, improving results on three multimodal memory benchmarks with about 90% fewer tokens than the leanest rival baselines.

Why it matters: For agents with long histories, how memories are retrieved and packed may drive cost more than the model does; these are benchmark results only so far.

From Drop #089 · Mon, Nov 16, 2026 · Morning
AgentsResearch

Preprint: scoped shared memory lifted agent pass rates with few privacy leaks

Researchers report in a preprint that MixUA turns agent runs into lessons scoped to a user, an agent, a pair or everyone. On their CrossLearnBench, pass rates rose up to 16 points over no memory, and at most 0.02% of privacy probes leaked.

Why it matters: If you plan shared agent memory across a team, decide who each stored lesson belongs to before it is written, and test whether private facts surface in other people's memory.

From Drop #087 · Sun, Nov 15, 2026 · Morning
AgentsResearch

Anthropic: stronger Claude models pick what to remember; dreaming corrects bad notes

In Anthropic's tests, Claude Fable 5 kept nine memory files and earned four Pokémon badges, while Sonnet 3.7 wrote 413 files and earned one. Rereading a session transcript let Claude catch and rewrite a wrong note.

Why it matters: Anthropic found fixed memory rules can hurt strong models. Check whether your agents' saved notes are ever tested against the record they came from.

From Drop #081 · Thu, Nov 12, 2026 · Morning
AgentsResearch

Preprint: agent memory works better when each lesson links back to its source run

Researchers report in a preprint that agents can fail tasks they would solve without memory by following wrong or misread lessons. Their WikiTrace ties each lesson to its source run and beat other memory methods on five benchmarks.

Why it matters: Lessons stored without evidence can mislead an agent. If you distill runs into notes, keep a link to the original trace so the agent can check before acting.

From Drop #081 · Thu, Nov 12, 2026 · Morning
AgentsResearch

Preprint: agents coordinated through a shared store nobody told them was shared

In a replicated study, four agents each saw a key-value store as private scratch space; joint success came only through it. One sentence telling agents to recheck the workspace switched coordination on and off.

Why it matters: Shared caches and memory between your agents act as a channel. The authors advise keeping that state separated, namespaced and logged.

From Drop #078 · Tue, Nov 10, 2026 · Evening
SecurityResearch

Preprint: every agent memory system tested could be poisoned across sessions

Researchers report in a preprint that MemoryDojo, 155 multi-session trajectories across five memory systems, found all of them susceptible to instructions planted in one session and acted on later, with wide variation in effects.

Why it matters: If your agents remember across sessions, treat stored memories as untrusted input and test what a planted note can make them do days later.

Preliminary Way in: malicious instructions saved to agent memory in one session, then retrieved and acted on later

Controls that would have helped:

From Drop #076 · Mon, Nov 9, 2026 · Evening
AgentsResearch

Anthropic: letting Claude judge what to remember beat fixed memory rules in its tests

Anthropic says a rule to record only failures cost Opus 5 and Opus 4.7 much of their gain from memory, and that dreaming, checking notes against session transcripts, fixed a wrong note in a Pokémon run.

Why it matters: Before writing detailed memory rules for your agents, try the simple version: tell the model what memory is for, and schedule a pass that checks its notes against the logs.

From Drop #071 · Sat, Nov 7, 2026 · Morning
AgentsResearch

Preprint: GUI agents struggle with vague requests that depend on a user's history

Researchers' MemU-Bench has 118 vague mobile tasks drawn from 29 long user histories; without that history, Gemini-3.1-Pro's SPL was 16.4%. Their AppMem memory layer lifted GPT-5.5 from 15.7% to 66.9% SPL.

Why it matters: For assistants that act on phones or desktops, test with the fuzzy, half-remembered requests people actually make; the authors report per-app structured memory beat a plain summary.

From Drop #063 · Tue, Nov 3, 2026 · Morning
AgentsResearch

Preprint: many dedicated agent memory systems underperform plain RAG

Researchers report that across multiple benchmarks and model backbones, many memory systems did worse than naive retrieval from raw history. They call the losses memory tax.

Why it matters: Before adding a memory layer, test it against simple retrieval; the authors say memory processing can remove or obscure evidence an agent needs.

From Drop #055 · Fri, Oct 30, 2026 · Morning
AgentsResearch

Preprint: many agent memory systems underperform plain RAG over raw history

Researchers evaluated memory systems across multiple benchmarks and model backbones and report many do worse than naive retrieval from raw history. They call the losses from reorganizing or distilling facts a memory tax.

Why it matters: Before adding a memory layer, compare it with simple retrieval: the authors say memory processing can remove or obscure the evidence an agent needs.

From Drop #050 · Tue, Oct 27, 2026 · Evening
AgentsResearch

Preprint: plain RAG over raw history beats complex agent memory in 6 of 7 settings

Researchers report Raw-RAG, dense retrieval plus reranking over raw history, ranked first in six of seven benchmark-model settings across four memory benchmarks, beating the strongest agentic memory system by 0.13-5.53 points.

Why it matters: The authors say preserving raw information matters and call naive RAG an essential baseline for the conversation tasks they tested.

From Drop #047 · Mon, Oct 26, 2026 · Morning
AgentsResearch

Preprint: judge agent memory by whether it improves future actions, not recall

Researchers propose Memory-Centric Continual Learning: a taxonomy of memory substrates, assets and policies, a test of retention under interference, adaptation and transfer, and a framework for how memory scales.

Why it matters: Most agent memory is scored as a retrieval cache, the authors argue; their protocol aims to separate real behavior gains from simply adding context.

From Drop #045 · Sun, Oct 25, 2026 · Morning
AgentsDocumented

LangChain details how Deep Agents offloads and summarizes context to fight context rot

LangChain says its open-source harness writes tool results over roughly 20,000 tokens to files, summarizes older messages at 85% of the context window, and keeps long-term memory in a plain file you control.

Why it matters: The ideas carry to any harness: offload rather than truncate, keep compaction recoverable, and own your memory so you can switch models.

From Drop #043 · Sat, Oct 24, 2026 · Morning
AgentsResearch

Preprint: PerMIT gates agent memory so untrusted entries need confirmation

Researchers report PerMIT gates untrusted, environment-derived memory, with integrity theorems machine-checked in Lean 4. They report zero observed attack success on Trojan Hippo and shared knowledge across up to 90 agents.

Why it matters: The authors call persistent memory a largely undefended surface for injection. Reducing attacks to one auditable confirmation is a design teams can study.

From Drop #040 · Thu, Oct 22, 2026 · Evening
AgentsDocumented

Databricks puts agent memory in Beta, built on its MemCurrent research

Databricks says Agent Memory keeps sessions, distills facts and preferences in a background pipeline, and runs on Lakebase. In its internal benchmarks the design tripled agent pass rate while using 31% fewer tokens.

Why it matters: Stateless agents redo yesterday's work; in Databricks' example a repeat question fell from 11 tool calls to 2, and developers can read, export and delete stored memories.

From Drop #037 · Wed, Oct 21, 2026 · Morning