Agent Memory
Synopsis
Agent memory is the layer that lets an agent carry the right state across a task instead of treating every prompt as a fresh start. In the World's Fair material, it shows up as several related patterns: long-context training and cache-augmented generation for keeping more source material available, semantic code retrieval for finding the right repository facts, context graphs and decision traces for remembering why choices were made, logs as durable agent state, and skill systems that package reusable procedures for future runs.
Origin And Context
This topic sits at the intersection of classic AI state management, knowledge representation, retrieval systems, observability, and database-backed application design. The connected sessions sharpen the tradeoff: Max Ryabinin frames long context as a training and systems problem, Luis Romero-Sevilla argues for knowledge representation when all context matters, Zach Blumenfeld pushes beyond documents toward decision traces and context graphs, and Victor Savkin's "A Genius With Amnesia" treats forgotten project context as a practical failure mode for software agents.
Why It Matters
Memory determines whether an agent can act consistently across multi-step work, codebases, users, and decisions. More context alone is not enough: Nupur Sharma's "Why More Context Makes Your Agent Dumber" and the Orbis cache-augmented generation talk both point to degradation when context is stuffed without structure. The stronger pattern is selective, source-grounded memory: retrieve the right facts, preserve provenance, keep decision history inspectable, and refresh or discard state when it becomes stale.
How To Use It
Separate working context from durable memory, then decide what each layer is allowed to retain. Store source-backed facts, repository architecture, user preferences, decision traces, logs, artifacts, and evaluation outcomes with timestamps and provenance. Use retrieval keyed to task intent, not just lexical similarity, and combine retrieval with graph or log structures when the agent needs relationships, chronology, or accountability. Add explicit policies for freshness, deletion, permissions, summarization, and OCR/transcript confidence, then test memory behavior with scenarios where stale or overbroad context would cause a wrong action.
Where It Is Useful
Memory is useful anywhere an agent has to continue work over time: coding agents that need repository history, customer-support agents that need prior case context, enterprise assistants that need decision records, research agents that need source trails, ETL remediation agents that need operational history, and personal productivity systems that must respect user preferences without relearning them every session.
When To Use It
Use durable memory when repeated interaction, long-horizon work, auditability, or cross-document reasoning matters. Prefer lighter working context for one-shot tasks. Avoid durable memory for sensitive data unless retention and deletion rules are clear, and avoid confident reuse of old state when the cost of asking again is lower than the risk of acting on stale context.
Active Use Cases
- Remembering repository architecture and prior implementation decisions.
- Maintaining user preferences and project constraints across sessions.
- Decision-trace retrieval for enterprise workflows.
- Long-context cache and knowledge-graph backed agent workflows.
Related Slide Decks
- youtube gcseUQJ6Gbg slides — Using OSS models to build AI apps with millions of users — Hassan El Mghari (4 extracted slide frames)
- youtube aHhB3sjGjkI slides — Agents Building Agents - Alfonso Graziano, Nearform (24 extracted slide frames)
- youtube jVjt 2g8NMY slides — A Genius With Amnesia - Victor Savkin, Nx (19 extracted slide frames)
- youtube EcqMYoIV57A slides — Why More Context Makes Your Agent Dumber and What to Do About It — Nupur Sharma, Qodo (4 extracted slide frames)
Related Scheduled Sessions
- 2026 06 29 anders swanson from context to memory your agents need a real memory layer — From Context to Memory: Your Agents Need a Real Memory Layer; Anders Swanson (Day 2 — Session Day 1 · 3:20pm-3:40pm · Expo Stage 2 NW; official schedule)
- 2026 06 30 stefania druga memory harnesses for long running research agents — Memory Harnesses for Long-Running Research Agents; Stefania Druga (Day 3 — Session Day 2 · 11:40am-12:00pm · Memory & Continual Learning; official schedule)
- 2026 06 30 prukalpa sankar wtf is the context layer the missing infrastructure for production agents — WTF Is the Context Layer? The Missing Infrastructure for Production Agents; Prukalpa Sankar (Day 3 — Session Day 2 · 1:55pm-2:15pm · Context Engineering; official schedule)
- 2026 06 30 elizabeth fuentes leone the infinite context window is a myth context engineering for ai agents — The Infinite Context Window Is a Myth: Context Engineering for AI Agents; Elizabeth Fuentes Leone (Day 3 — Session Day 2 · 3:20pm-3:40pm · Expo Stage 3 SW; official schedule)
- 2026 07 01 stephen chin crabrag why automated assistants need graph memory not more tokens — CrabRAG: Why Automated Assistants Need Graph Memory, Not More Tokens; Stephen Chin (Day 4 — Session Day 3 · 10:45am-11:05am · Graphs; official schedule)
- 2026 07 01 james le video has no memory here s how we built one — Video Has No Memory. Here's How We Built One.; James Le (Day 4 — Session Day 3 · 2:25pm-2:45pm · Graphs; official schedule)
- 2026 06 29 ignacio martinez total recall agent memory and harness engineering — Total Recall: Agent Memory and Harness Engineering; Ignacio Martinez (Day 1 — Workshop Day · 9:00am-11:00am · Workshops Day 1; official schedule)
- 2026 06 29 louis fran ois bouchard context engineering in 2026 compaction memory and cost — Context Engineering in 2026: Compaction, Memory & Cost; Louis-François Bouchard, Samridhi Vaid, Omar Solano (Day 1 — Workshop Day · 2:20pm-4:20pm · Track 6; official schedule)
- 2026 06 30 anant srivastava prompt memory weights the architecture decisions most ai teams make by accident — Prompt, Memory, Weights: The Architecture Decisions Most AI Teams Make by Accident; Anant Srivastava (Day 3 — Session Day 2 · 12:05pm-12:25pm · Context Engineering; official schedule)
- 2026 06 30 shlok khemani lessons from studying every memory system — Lessons from Studying Every Memory System; Shlok Khemani (Day 3 — Session Day 2 · 3:20pm-3:40pm · Memory & Continual Learning; official schedule)
- 2026 06 29 yoni michael the data context layer why data engineering agents need more than code and databases — The Data Context Layer: Why Data Engineering Agents Need More Than Code and Databases; Yoni Michael, Brandon Callender (Day 1 — Workshop Day · 2:20pm-4:20pm · Track 2; official schedule)
- 2026 06 30 jack morris scaling compute on context — Scaling Compute on Context; Jack Morris (Day 3 — Session Day 2 · 11:40am-12:00pm · Memory & Continual Learning; official schedule)
- 2026 06 30 brandon waselnuk your agents lack context here s how to fix you re absolutely right — Your agents lack context: Here's how to fix "You're absolutely right!"; Brandon Waselnuk (Day 3 — Session Day 2 · 12:05pm-12:25pm · Context Engineering; official schedule)
- 2026 06 30 gil feig why your company needs a context graph and how to build it — Why your company needs a context graph, and how to build it; Gil Feig (Day 3 — Session Day 2 · 1:55pm-2:15pm · Expo Stage 3; official schedule)
- 2026 07 01 omri bruchim from systems of record to systems of context — From Systems of Record to Systems of Context; Omri Bruchim (Day 4 — Session Day 3 · 12:05pm-12:25pm · Graphs; official schedule)
- 2026 07 01 yuchen fama kv cache aware routing and p d disaggregation on kubernetes the parts public benchmarks don t show — KV Cache-Aware Routing and P/D Disaggregation on Kubernetes: The Parts Public Benchmarks Don't Show; Yuchen Fama, Ashish Kamra (Day 4 — Session Day 3 · 2:50pm-3:10pm · Inference; official schedule)
- 2026 06 30 rishab kumar from stateless to stateful orchestrating real time voice and messaging agents with twilio and amazon bedrock — From Stateless to Stateful: Orchestrating Real-Time Voice & Messaging Agents with Twilio and Amazon Bedrock; Rishab Kumar (Day 3 — Session Day 2 · 12:05pm-12:25pm · Expo Stage 2 NW; official schedule)
- 2026 06 30 omer primor the rise of caas context as a service for agentic ai — The Rise of CaaS: Context-as-a-Service for Agentic AI; Omer Primor (Day 3 — Session Day 2 · 1:55pm-2:15pm · Computer Use; official schedule)
- 2026 06 30 rachna srivastava guardians of the state how we built an air gapped ai fortress for consumer data — Guardians of the State: How We Built an Air-Gapped AI Fortress for Consumer Data; Rachna Srivastava (Day 3 — Session Day 2 · 1:55pm-2:15pm · AI-Native Enterprises; official schedule)
- 2026 07 01 brandon waselnuk beyond rag see a relational context engine reduce token burn — Beyond RAG: See a relational context engine reduce token burn; Brandon Waselnuk (Day 4 — Session Day 3 · 11:10am-11:30am · Expo Stage 1 NE; official schedule)
- 2026 07 01 kay malcolm no memory no harness why the database is the last line of defense — No Memory, No Harness: Why the Database Is the Last Line of Defense; Kay Malcolm (Day 4 — Session Day 3 · 2:50pm-3:10pm · Harness Engineering; official schedule)
- 2026 06 30 peter werry how to generate mergeable code with a context engine — How to generate mergeable code with a context engine; Peter Werry (Day 3 — Session Day 2 · 11:40am-12:00pm · Expo Stage 2 NW; official schedule)
- 2026 06 29 krishna prasad srinivasan from scratch to sota training a 3b state space vision model for 1 4 billion people — From Scratch to SOTA: Training a 3B State-Space Vision Model for 1.4 Billion People; Krishna Prasad Srinivasan (Day 2 — Session Day 1 · 3:20pm-3:40pm · Vision & OCR; official schedule)
- 2026 07 01 karthik ranganathan agent memory is a solved problem agent learning is not — Agent Memory Is a Solved Problem. Agent Learning Is Not.; Karthik Ranganathan, Heather Downing (Day 4 — Session Day 3 · 3:20pm-3:40pm · Expo Stage 1 NE; official schedule)
Related People
- John Lindquist
- Brandon Waselnuk
- Peter Werry
- Joseph Nelson
- Ahmad Osman
- Ido Salomon
- Yuval Belfer
- Harshul Jain
- Tanmay Sah
- Christopher Manning
- Merve Noyan
- Anders Swanson
- Stefania Druga
- Prukalpa Sankar
- Elizabeth Fuentes Leone
- Stephen Chin
- James Le
- Ignacio Martinez
- Louis-François Bouchard
- Samridhi Vaid
- Omar Solano
- Anant Srivastava
- Shlok Khemani
- Yoni Michael
Related Companies
- Together AI
- Microsoft
- Unblocked
- Neo4j
- NVIDIA
- OpenAI
- Oracle
- Anthropic
- Towards AI
- MCP Apps
- AI21
- egghead.io
- typedef
- Red Hat
- Yugabyte
- Roboflow
- Meta
Transcript And Resource Support
Transcript-backed resources
- youtube EcqMYoIV57A — Why More Context Makes Your Agent Dumber and What to Do About It — Nupur Sharma, Qodo
- youtube B9h9ovW5H9U — Why your agents need decision traces, not just documents — Zach Blumenfeld, Neo4j
- youtube TUnPNY4E2fw — Road to 5 Million Tokens: Breaking Barriers in Long Context Training — Max Ryabinin, Together AI
- youtube UNzCG3lw6O0 — Building Great Agent Skills: The Missing Manual
- youtube zKk7sDMGDEQ — Benchmarking semantic code retrieval on Claude Code — Kuba Rogut, Turbopuffer
- youtube XovaGv4f39A — When All Context Matters: Extended Cache Augmented Generation - Luis Romero-Sevilla, Orbis
- youtube jVjt 2g8NMY — A Genius With Amnesia - Victor Savkin, Nx
- youtube UPwGaM2MKHY — The Log Is The Agent - Ishaan Sehgal, Omnara
- youtube LrGCT7G_rU8 — Using RL Agent to Detect and Remediate ETL Pipeline Failures - Anna Marie Benzon
- youtube Jx4ZFEAq6bY — User Signal Dies at the Retrieval Boundary - Sonam Pankaj, StarlightSearch
- youtube r305 aQTaU0 — Text Diffusion — Brendan O’Donoghue, Google DeepMind
- youtube spNAUEgq_A8 — The Future Is Domain-Specific Agents - Justin Schroeder, StandardAgents
- youtube IJXjTLPzvAU — The Miranda Hypothesis: How Hamilton Poisoned Persona Evals - Jacob E. Thomas, Results Gen
- youtube sAOBXCDiDOs — MCP Apps: Primitives, discovery, and the Future of Software - Pietro Zullo, Manufact, Inc
- youtube YYH0DMQr30A — Task Fidelity Scaling Laws — Kobie Crawdord, Snorkel
- youtube htM02KMNZnk — WF2026: Software Factories & Keynotes ft. Microsoft, OpenAI, OpenClaw, Z.ai (GLM), MiniMax, HF
- youtube ZD9 4fW2HhM — Build Systems, Not Code - Angie Jones, Agentic AI Foundation
- youtube 3hXJI2q0Jz8 — Recursive Coding Agents - Raymond Weitekamp, OpenProse
Quote signals
- “And today, I'm going to tell you about our research project, which is called Road to 5 million sequence length, breaking memory barriers in context parallelism.” — youtube TUnPNY4E2fw
- “Um to do that all effectively, you need to make sure that the models are able to process that context and work with it correctly at the training time.” — youtube TUnPNY4E2fw
- “This approach would look something like "cache augmented generation" (CAG), where we use a model with a large context window, load the documents into the context, and cache the context by storing the model's KB matrix.” — youtube XovaGv4f39A
- “Uh as you continue scaling your context, your memory keeps growing linearly, which is not as bad, but still pretty difficult to deal with, unless you apply a range of specific techniques.” — youtube TUnPNY4E2fw
- “I'm on a mission to solve knowledge representation when all context matters.” — youtube XovaGv4f39A
- “And so, to kind of think actually about what a context graph is, we need to ask ourselves, "Would you agents really be accurate?" Right?” — youtube B9h9ovW5H9U
- “And then context, policies that are um in different reasoning by AI that records memory, but um by employees and and past humans that have made decisions.” — youtube B9h9ovW5H9U
- “Um But yeah, that's a good point and then in the create context graph, we're still working on how you would write um new decision traces.” — youtube B9h9ovW5H9U