For long-running agent work, the most useful change is usually not a bigger prompt. It is keeping selected information outside the active context, then retrieving only the pieces a given step needs. That pattern is supported by published designs and evaluations, but the results are specific to the systems and tasks tested, and a memory layer adds its own failure modes. This article explains what each approach does, where the reported evidence stops, and how to decide between them.
Why carrying raw history stops working
An agent that runs for hours or days accumulates tool outputs, failed attempts, intermediate drafts, and user corrections. The simplest design appends all of it to every new request. That works until the history outgrows the window, and it becomes expensive and noisy well before then. Most of the history is irrelevant to the current step, and the parts that matter, such as a decision made three hours earlier or a dependency between two files, are buried among material that does not.
As an Amazon Associate I earn from qualifying purchases.
A larger context window raises the ceiling but does not decide what is useful. Putting more text in front of a model does not tell it which fact governs the current task. That gap is the reason memory systems are being built as a separate layer.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsContext, compaction, and durable memory are different things
These three terms are often used interchangeably, and they solve different problems.
#1 Best Overall
- Context is everything the model can use during one inference step. It is the working input, nothing more.
- Compaction summarizes a running session near a context limit so the work can continue in a fresh window. Anthropic describes keeping critical decisions and unresolved work while dropping redundant content, and it warns that aggressive compaction can discard details whose importance only emerges later.
- Durable memory stores information outside the prompt, so it survives beyond the current session. Relevant entries must be retrieved and placed into context before the model can use them.
The approaches can be combined. A session can be compacted for immediate continuity while notes or structured records carry what must survive the restart.
The main approaches compared
| Approach | What is kept | How it returns to the model | Main risk |
|---|---|---|---|
| Larger context window | Full raw history, until the limit | Always present, if it fits | Cost and noise grow; no judgment about relevance |
| Compaction | A summary of the running session | Summary replaces earlier turns | Details dropped in summarization may matter later |
| Structured note-taking (agentic memory) | Agent-written notes on progress, decisions, dependencies | Agent reads notes back in when needed | Notes can go stale or be incomplete |
| Knowledge-centric memory | Facts or reusable skills derived from interactions | Retrieval selects task-relevant units | Wrong or missed retrieval; update and correction rules are harder |
The table describes design patterns, not rankings. None of these sources establishes that one approach is best across all agent workloads.
A simple persistent-notes pattern
Anthropic’s engineering article describes structured note-taking, or agentic memory, as a technique in which the agent regularly writes notes persisted to memory outside of the context window. It is the simplest form of durable memory to start with, because it needs no retrieval index. A minimal version looks like this:
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →- Define a notes file or store with fixed sections, such as goal, decisions made, open items, and dependencies.
- At natural checkpoints (a finished subtask, a changed decision, a discovered constraint), have the agent write a short entry to the matching section.
- At the start of a new session, load the notes file into context before the first task step.
- When the agent finds a note is outdated, have it revise or mark the entry rather than appending a contradiction.
- Periodically check the notes against the actual state of the work, such as the repository or the ticket, and correct any drift.
The step most often skipped is the last one. Notes are only as reliable as the update discipline behind them.
Rank #2
- Capture Every Milestone from Birth to Age 5: From birth to age 5, this complete baby memory book includes 128 guided pages to help you document every milestone. The simple, organized layout makes it easy for busy parents to fill out this first year memory book without feeling overwhelmed
- 6 Keepsake Envelopes for Precious Mementos: Unlike other books, ours includes 6 built-in envelopes to safely store physical memories. Store hospital bracelets, ultrasound photos, first haircut locks, and special cards all in one organized place
- From Pregnancy to First Year Memories: Capture your journey from the pregnancy story and gender reveal to the baby's arrival and family tree. This baby milestone book includes space for footprints and many other meaningful moments that become cherished memories for a lifetime
- 24 Free Milestone Stickers Included: Celebrate your baby's growth with a set of 24 milestone stickers for monthly photos and special celebrations. This added value makes our baby book a standout choice for tracking your little one's progress through their early years
- Gift-Ready Keepsake Box for Baby Registry: Presented in a premium sliding gift box with gold foil details, this book makes a beautiful baby shower gift or baby registry essential. A thoughtful Mother's Day gift for new moms who value quality and style
Structured and knowledge-centric memory
Notes work well when one agent tracks one project. Other designs try to reuse what was learned across many interactions.
Knowledge-centric memory (PlugMem)
Microsoft Research describes PlugMem as converting interactions into structured facts or reusable skills, then retrieving and distilling the knowledge relevant to the current task. The authors, Ke Yang, Michel Galley, Chenglong Wang, Jianfeng Gao, and academic collaborators, frame the problem this way: “It seems counterintuitive: giving AI agents more memory can make them less effective.” Their argument is that the fix lies in organizing memory into reusable knowledge and retrieving task-relevant units, not in storing more raw material. Microsoft Research reports that PlugMem outperformed generic retrieval methods and task-specific memory designs across three benchmarks while using significantly less memory-token budget. The published page does not give a numeric improvement, so no figure should be quoted from it.
Gist memory with lookup (ReadAgent)
ReadAgent, from Google DeepMind researchers in 2024, partitions a long document into episodes, writes a concise gist for each, and retrieves the original passage when more detail is needed. The useful idea is pairing compression with access to source text. A pure summary is lossy; a summary that can return to the original keeps the route to the evidence open.
Free tools Windows power users keep installed
One-click scans. No signup required.
What the reported results do and do not show
- ReadAgent’s reported extension of effective context length by 3–20× comes from evaluations on QuALITY, NarrativeQA, and QMSum. It describes those tasks and that system, not agents in general.
- PlugMem’s reported gains are benchmark results from Microsoft Research, a vendor-affiliated source. They should be read as the authors’ findings on three named benchmarks.
- The AAAI Symposium Series review identifies separating memory types and managing memory over an agent’s lifetime as open problems, and notes that vector databases are a common implementation for long-term memory. A vector database is one storage choice, not a requirement.
- The 2026 AMA-Bench paper argues that dialogue-only memory evaluations miss continuous agent-environment trajectories, and reports weaknesses in capturing causal and objective information through similarity-based retrieval. That is the paper’s finding, not a settled field-wide conclusion.
Taken together, the evidence supports selective storage, organization, and retrieval as sound design patterns. It does not show that memory reliably improves every agent.
Rank #3
Where memory goes wrong
- Stale facts. A stored decision is no longer true, but the agent still retrieves it and acts on it.
- Contradictory entries. Two notes disagree because one was never updated when the other was written.
- Missed retrieval. The relevant entry exists but is not surfaced for the task at hand.
- Irrelevant recall. A retrieved entry is similar in wording but wrong for the situation, which can mislead the agent more than no memory at all.
- Over-compression. A summary drops a constraint or a causal link that only matters later.
- Privacy exposure. Persisted notes can contain user data that should not be kept or shared. The sources reviewed raise privacy as an open question without giving a standard solution.
Retrieved memory enters the context when it is used, so it competes for the same space as everything else. Storing more does not guarantee the right entry arrives at the right step.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to evaluate a memory setup
Test the system on the work it has to do, not on a similarity score. Check these points:
- Does it preserve goals, decisions, dependencies, and causal information, or only similar text?
- Can it return the original supporting passage with its source context?
- How are stale or conflicting entries corrected or removed?
- What happens when the right memory is not retrieved?
- What does each step cost in tokens and latency, including retrieval overhead?
- Do the evaluation tasks resemble the agent’s real actions over a long horizon?
If a benchmark measures only conversational recall, it will not answer these questions for an agent that edits files or calls tools.
Which approach to start with
Use these rules as a starting point, not a verdict:
- If the job fits comfortably in the window and ends in one session, a larger context is simpler and may be enough.
- If the job spans sessions but has a clear project state, start with a structured notes file and a disciplined update step.
- If the agent must reuse knowledge across many tasks or documents, evaluate knowledge-centric retrieval, with a test set that resembles your own workload.
- In most long-running systems, combine them: compact the live session, and keep durable notes for what must survive.
Anthropic’s developer platform offers a file-based memory tool, as the engineering article describes. Check the current documentation for availability, terms, and region before building around it.
Frequently Asked Questions
Do I need a vector database to give an agent memory?
No. The AAAI Symposium Series review notes that vector databases are a common implementation for long-term memory, but a notes file that the agent writes and reloads at the start of each session is also a form of persistent memory and needs no retrieval index.
Is a memory system a replacement for a larger context window?
No. The two solve related but different problems. The context window is the working input for one step, while memory persists selected information outside it and must retrieve the relevant parts when they are needed.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Quick Recap
The Bottom Line
“”
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
