Sensory
Recent perceptual traces, such as visual observations retained for a physical task.
Example: a robot tracks an object briefly hidden by its arm.
Research synthesis · Checked 7 September 2026
The Architecture of Agent Memory
What an agent keeps, how it changes, and whether it helps next time. Current research reaches beyond recall into evolving facts, reusable experience, and the work of maintaining memory. [02]
MutableFacts can be revised
SelectiveContext has a budget
MultimodalEvidence has many forms
TraceableOrigins matter
Themes in the sources below, rather than measured shares of the field.
01 / Organize the vocabulary
Memory Taxonomy
Purpose, representation, and scope describe different aspects of the same system. A vector index is a retrieval structure; episodic memory describes the kind of information retained. [01] [03]
What is remembered?
Functional vocabulary helps describe what a memory is for. One system can support several of these at once.
Working
The information currently available for the next decision or response.
Example: the current request, a plan, and selected evidence.
Semantic
Facts and concepts the system can reuse.
Example: a collaborator’s stated preference or a project decision.
Episodic
Records of particular encounters, actions, or outcomes.
Example: the steps attempted in yesterday’s failed task.
Procedural
Reusable ways of doing something, represented in instructions, code, skills, or learned behavior.
Example: a procedure improved after an earlier mistake.
02 / Follow the work
The Memory Loop
A useful architecture makes revision and feedback part of the path. Select a stage to see the decision it contains. [19] [13]
01 · Observe something
Encounter
Return path Feedback may change the record, the retrieval rule, or the next action. Keeping an event and learning from it are separate achievements. [15]
03 / Compare implementations
Techniques & Tools
Selected implementations, with research systems labelled separately. Documentation describes capabilities; the tradeoffs below are this exhibit’s synthesis, not a performance ranking.
| System | Focus | What it offers | What to consider |
|---|---|---|---|
| LangGraphApplication tooling | Thread checkpoints and cross-thread stores | Explicit application control over state and when memory is read or written. | You still design selection, updates, and the backing store. |
| LettaApplication tooling | Editable memory and background “dreaming” | Current docs describe a git-backed memory filesystem with consolidation between active work. | Background revisions consume compute and need review for drift. |
| Mem0Application tooling | Extracted memories and multi-signal search | APIs for adding, searching, correcting, and deleting memories; managed and self-hosted paths. | Current docs distinguish Platform graph features from OSS. Additive extraction does not automatically replace an old fact. |
| ZepApplication tooling | Facts with relationships and validity times | Records when a fact applies and when the system learned or invalidated it. | Extracted relationships and inferred validity still need evaluation on the application’s data. |
| MEMRobotics research | Memory for vision-language-action policies | Combines compressed short-horizon video with longer-horizon text memory for robotic tasks. | A task-specific research architecture; deployment evidence here concerns robot policies. |
04 / Ask what was tested
Benchmarks & Blind Spots
Multimodal and agent-action benchmarks already exist. The harder question is which conditions each one tests, and how well its results travel to your use case. [11] [10]
8 benchmarks · choose one to examine its scope
LongMemEval
Can a chat assistant recall information, reason across sessions and time, handle updates, and abstain?
Scope: Conversational questions do not establish performance in physical or tool-driven tasks.
MemoryAgentBench
Can memory support retrieval, test-time learning, long-range understanding, and selective forgetting as information arrives?
Scope: Its incremental datasets cover selected competencies rather than every deployment condition.
AMA-Bench
Can an agent use long histories of agent-environment interaction, including machine-generated traces?
Scope: Real and synthetic trajectories probe specific tasks; scores are not a universal memory rating.
M³Eval
What survives video-memory interference, and can a model track the source of what it remembers?
Scope: Designed for video memory. It is an evaluation suite, not an agent memory backend.
M³Exam
Can an assistant connect text, images, and documents across sessions, including implied context?
Scope: Tests multimodal conversational reasoning in its constructed interaction scenarios.
StateMemBench
Does an answer use the current state, a superseded state, or neither?
Scope: A new August preprint with a particular scenario and grading design; wider transfer needs testing.
StreamMemBench
Does evidence from an observed stream and feedback from one task help with a subsequent task?
Scope: The August revision separates evidence use, feedback incorporation, and later reuse.
LongMemEval-V2
Can memory retain web-environment states, workflows, recurring gotchas, and awareness of a question’s premise?
Scope: A separate environment-experience benchmark; its scores cannot be exchanged with the original LongMemEval’s.
Before comparing scores: match the dataset version, base model, memory budget, and grading method. A score for conversational recall does not tell you how a robot will behave. Cost and update reliability also belong in the comparison. [18] [07]
05 / Since the July synthesis
Recent Evidence
Selected work published or revised after 17 July 2026. These are research proposals with bounded evaluations; the implications are our reading of their design lessons.
Compression has a design problem
RSM-full studies how to merge incoming memories and pack evidence into a small prompt budget. The reported gains are specific to that constrained setting.
Design implication: Compare answer quality alongside the amount of context used.
[18]Several views of the same history
Agent Zero Memory proposes an event timeline, an entity-event graph, and a documentary store, linked to evidence. Its evaluation reports results on conversational benchmarks.
Design implication: Source tracing and complementary representations are useful design questions; the paper’s claims still need independent testing.
[17]Remembering what is current
StateMemBench distinguishes a current answer from one based on an obsolete state. StateMem explicitly represents supersession and dependencies.
Design implication: Test revisions directly instead of assuming good recall handles them.
[14]Learning what to keep
AttriMem adds localized feedback to reinforcement learning for memory construction, connecting intermediate memory choices to their contribution to an answer.
Design implication: Memory management itself can be a learned policy, alongside hand-written rules.
[16]06 / What remains difficult
Open Questions
When does a memory stop being current?
History can remain useful after a fact changes. A system needs to distinguish superseded information from the version that should guide the next answer. [14]
Does feedback improve the next task?
StreamMemBench finds a gap between retaining evidence or feedback and using it reliably afterward. Evaluation needs to follow that whole path. [15]
What survives compression?
Compact summaries and indexes make long histories tractable, but the selection and assembly of evidence influence what remains answerable. [18]
Who can revise shared memory?
Persistent state also carries scope, permissions, and commitments. Correction, provenance, and recovery matter when multiple participants depend on the same record. [19]
07 / Follow the sources
References & Update Notes
Checked 7 September 2026. Paper dates distinguish first publication from later revisions where relevant. Product documentation can change after this check.
Primary sources used in this exhibit 20 references
- Memory in the Age of AI Agents
Distinguishes the form, function, and changing operations of memory.
- A Survey of Agent Memory in the Second Half
Connects memory substrates, cognitive mechanisms, and whose experience is retained.
- LangChain / LangGraph: Memory overview
Thread state, cross-thread stores, memory types, and foreground/background writing.
- Letta: Memory & dreaming
Editable MemFS memory and background consolidation.
- Mem0: How Mem0 Works
Extraction, multi-signal search, explicit corrections, and Platform/OSS differences.
- Zep: Facts
Fact validity intervals, updates, and the difference between event time and recording time.
- MEM: Multi-Scale Embodied Memory for Vision Language Action Models
Robot policies combining short-horizon video with longer-horizon language memory.
- LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory
Extraction, reasoning across sessions and time, knowledge updates, and abstention.
- MemoryAgentBench: Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions
Retrieval, test-time learning, long-range understanding, and selective forgetting.
- AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications
Real and synthetic agent-environment trajectories.
- M³Eval: Multi-Modal Memory Evaluation through Cognitively-Grounded Video Tasks
Video memory probes including interference and source tracking.
- M³Exam: Benchmarking Multimodal Memory for Realistic User-Agent Interactions
Multi-session text, image, and document memory with cross-modal reasoning.
- WorldMemArena: Evaluating Multimodal Agent Memory Through Action-World Interaction
Examines memory writing, maintenance, retrieval, and use in evolving multimodal tasks.
- Can Agent Memory Systems Track Evolving State?
StateMemBench and StateMem distinguish current facts from superseded ones.
- StreamMemBench: Streaming Evaluation of Agent Memory for Future-Oriented Assistance
Tests whether observed evidence and feedback help a later task.
- AttriMem: Attribution-Guided Process Feedback for Agent Memory Construction
Learns memory-construction decisions with more localized process feedback.
- Agent Zero Memory: Provenance-Aware Long-Term Memory for LLM Agents
Combines an event timeline, an entity-event graph, and a documentary memory with source pointers.
- Compact-Memory LLM Agents via Online Max-Member Clustering and Atom-Aware Packing
Studies how merging memories and packing retrieved context affect quality under tight budgets.
- Always-On Agents: A Survey of Persistent Memory, State, and Governance in LLM Agents
Extends the view to permissions, commitments, provenance, recovery, and externally committed effects.
- LongMemEval-V2: Evaluating Long-Term Agent Memory Toward Experienced Colleagues
Tests retained experience of web environments, including workflows and recurring failure modes.
What changed from the supplied July dashboard
- Removed the unsourced 75/15/10 research-focus chart. Its replacement is an explicitly illustrative state-tracking example, with no invented measurements.
- Replaced blanket absence claims with named multimodal and agent-action benchmarks. Some were already available before the July dashboard; this update corrects that coverage as well as adding newer work.
- Separated memory purpose, representation, and scope. Episodic memory is no longer grouped as though it necessarily requires embodiment.
- Separated frameworks and research systems from evaluation suites, and replaced the single unavailable synthesis citation with direct primary sources.
- Preserved the supplied Sora/IBM Plex typography, dark panel layout, warm palette, and base CSS. Added restrained cyan highlights and functional selectors. Removed the duplicated HTML wrapper and runtime chart/style dependencies.
The original synthesis report named by the upload was not supplied. This update checks the dashboard’s visible claims against the linked papers and documentation. It is a selected research update, not an exhaustive literature review or an independent replication.