Agent Memory

DolphinBench tests whether agent memory improves real task completion, while requiring accuracy, cost and latency so retrieval quality cannot hide impractical tradeoffs.

Repeated peer verification can drift into collusion when protocol compliance conflicts with rewards; limiting shared history reduced the effect in this experiment.

V7 uses GPT-5.6 to turn scattered company files into source-linked context for agents, targeting complex work that needs institutional knowledge.

MDL adds a white-box trust decision between retrieval and generation, using relevance, reliability, and task risk to reject conflicting memories without another model call.

Bayesian Chronicle Agents separate explicit belief state from generated speech, making persuasion strength configurable and multi-agent opinion changes auditable.

Treat memory, context control, concurrency, and model routing as harness concerns. The workshop shows how these layers can improve agent behavior without changing model weights.

Pinecone’s Nexus packages persistent company context behind an agent-facing query layer. The useful idea is to maintain goals, decisions, tables, and indexes once instead of rediscovering them per session.

Mind2Dialogue trains assistants from simulated, evolving user states that are hidden at deployment. The approach could inform persistent agents, but its reported gains concern personalization rather than coding work.

Oracle’s team used a shared memory broker to carry agent context across people and sessions. The useful pattern is durable, queryable team memory; the database choice is vendor-framed.

A production agent's harness is everything around the model: tools, memory, runtime, isolation, and deployment controls. The practical lesson is to standardize that layer across agents.

ConvMem replaces sequential long-context memory updates with a parallel hierarchy of summaries, offering a training-free pattern for agents that must reason over oversized inputs.

Procedural Graphs externalize an agent’s workflow, guide each action from the relevant subgraph, and refine the structure from failed and working runs. This targets drift and repeated tool mistakes on long tasks.

MeClear suppresses memories that hurt the current task, using cooperative attribution to expose conflicts that single-memory removal can miss, while leaving the persistent memory bank unchanged.

Agent memory can break across model upgrades even when the store is unchanged. Test migrations by direction, rebuild embeddings completely, and retain raw histories for repair.

This survey separates verbal feedback by when it acts and what it changes, helping agent builders distinguish task specification, test-time guidance, and parameter updates.

S3Gym tests whether agents can explore, judge outcomes, and reuse experience. Raw history, compressed memory, and training each help in some tasks, but none is a reliable default.

Exa’s internal agent clone combines writing examples, decision evals, and permission-scoped tools—a concrete pattern for encoding judgment without sharing full access.

Reusable agent skills transferred better when induced from subtasks and stored as text. Whole-task skills often hurt versus no memory, so retrieval alone is not evidence that a memory helps.

Compound engineering treats each agent mistake and product judgment as reusable context. The useful discipline is to invest in memory and review so the next feature becomes easier, not merely faster.

BATON explores robot subtasks independently, stores their solutions, and manages handoffs between them, replacing multiplicative whole-task search with additive composition.

Agent handoffs should preserve task-relevant predictive state, not blindly compress the transcript. The paper offers a three-part record for carrying decisions, evidence, and irreducible observations across sessions.

SheetCompass keeps spreadsheet structure as a hierarchical relation graph instead of flattening cells into text. Its agent memory retains task-relevant context across sheets, tables, and columns.

Persistent memory is a compute and product tradeoff, not just retrieval. Profiles need conflict detection, visibility, editing, and deliberate update cadence before agents can rely on them.

A useful agent knowledge base starts with abundant raw Markdown, then adds controlled tags, backlinks, and generated indexes. Scheduled agents can maintain it, but sync and taxonomy need explicit rules.

Stronger models can still behave like smart novices inside company-specific workflows. Agents need learning loops that compress experience into reusable context, procedures, and judgment without losing reliability.

Training directly on a private corpus can memorize it without producing useful behavior. Durable context learning likely needs synthetic study tasks and recursive self-improvement, not a single fine-tuning pass.

Always-on production agents can turn releases and operational events into tailored monitoring tasks. The useful pattern is persistent context plus scoped triggers, checks, and escalation.

SkillProx evolves reusable agent instructions through measured retries, regression rollback, and utility-based pruning, reporting a 3-point accuracy gain over its strongest baseline.

Blast Radius predicts which code and context a prompt will reach, then reversibly archives unused history. Across seven OpenAI models, it cut token consumption by 17–26%.

PsychoAgent separates factual and affective memory, then reranks relevant memories by salience. It retrieved more conflict-critical context, but output-quality differences were not significant.

Chained RLM repeatedly gives the same model a fresh context while preserving summaries, a blackboard, and artifacts—a concrete harness pattern for limiting error buildup.

PRECOG precomputes SSM corpus states and injects them at query time, avoiding context re-ingestion. On a 1.2B edge model, reported prefill fell from about 27 seconds to under 6 ms.

This survey organizes long-horizon agent weaknesses into five capability gaps, offering a useful checklist for harness design and evaluation rather than a new implementation.

Shared agents need a different harness: action-boundary security, selective group memory, per-user privacy, and routing that decides both who receives information and when the agent should speak.

UniMem routes novel experience into retrieval memory and consolidates recurring patterns into expandable parameters, offering a design for agents that learn across unlabelled task streams.

MemLens scores individual agent memories, stores them by value, and exposes quality, latency, and token tradeoffs—an inspectable alternative to retaining every interaction equally.

Netflix’s performance-agent playbook feeds production profiles and exact deployed code to agents, then relies on canaries and engineers to validate proposed optimizations.

A delayed KV-eviction policy helps when memory reuse is sharp and observable, but loses its edge on natural-text benchmarks—a useful warning for agent-memory claims.

LLM synthesis breaks simple source pointers when facts merge, change, or outlive one input. Model provenance inside the context graph so agents can filter trust, debug derivation, and apply deletion policies.

Atlan’s agent experiments argue for shared, versioned context instead of per-agent memory: a portable layer for business facts, skills, norms, retrieval, and feedback across changing harnesses.