Free tools Windows power users keep installed
One-click scans. No signup required.
A larger context window gives an AI agent more room to read; it does not ensure that the agent can tell which facts are current, which have been superseded, or which still conflict. Somay’s PaperPlanes project takes a different approach: retain knowledge over time, preserve conflicting claims as disputed, and let the system inspect earlier database state. The author describes this as a bi-temporal memory engine, but the project write-up does not provide enough schema detail to independently assess that temporal model.
Why a million-token context is not the same as memory
Context capacity and memory quality solve different problems. A long prompt can include more documents or conversation history, but that material may contain outdated information alongside newer claims. More input alone does not tell an agent which version to trust or how a fact changed.
As an Amazon Associate I earn from qualifying purchases.
That distinction is the central argument of Somay’s Aug. 18, 2026, DEV Community post about PaperPlanes, a stateful research assistant. The author’s point is not a measured finding that a particular context-window size fails; it is an architectural warning that expanding the input does not, by itself, resolve conflicting or stale knowledge.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?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 & 11Instead of treating memory as a bigger prompt, PaperPlanes treats it as persistent state: facts are retained, conflicts are surfaced, and earlier database states can be inspected. Somay summarizes the philosophy as: “Memory has to be the product itself.”
#1 Best Overall
How PaperPlanes handles facts that change
Keep competing claims instead of silently overwriting
According to Somay, the system does not delete an older fact when new information contradicts it. It retains both and marks them disputed. This makes a disagreement visible to later retrieval rather than concealing it behind whichever value happened to be written last.
That policy is useful when a fact can genuinely change or when sources disagree: the system can preserve the conflict for review rather than presenting an unqualified answer. The post does not specify a complete conflict-resolution process, such as how a user or agent decides which claim should eventually be treated as current.
Inspect prior database state
The author says CockroachDB’s AS OF SYSTEM TIME queries support a UI slider for viewing what the system believed at an earlier time. This is a way to inspect historical database state. It should not automatically be read as proof that every fact has a fully defined valid-time history: the post does not provide enough schema detail to assess how the claimed bi-temporal model represents when a fact was true versus when the system recorded it.
What the reported implementation includes
Somay describes PaperPlanes as a stateful research assistant built with CockroachDB for storage, AWS Bedrock using Amazon Nova Pro and Titan Embeddings, and LangGraph with FastAPI for agent logic. These are the author’s reported technology choices, not a comparative finding that this stack is best for agent memory.
Agent access to its own memory system
The post also says an Amazon Nova agent uses a CockroachDB Managed MCP Server tool to run read-only SQL queries about the database schema, memory state, and metadata. In the described design, this gives the agent a way to inspect its stored state. The write-up does not independently establish how broadly that capability is used or how it performs in other deployments.
A time-based retrieval heuristic
Somay gives the decay expression R = e^(-t/S), where retention R declines as time t passes unless access reinforces memory strength S. The author presents this as an implemented heuristic. The post does not state calibrated units, parameter-fitting methods, or evaluation results, so the expression should not be treated as a validated general law of agent memory.
What the concurrency test does—and does not—show
Somay reports a project-specific test in which 25 concurrent agent writers used one memory node. The author says the run detected 58 transaction collisions, retried them automatically, and lost 0 writes. Those figures describe that reported test, not a general performance or reliability guarantee.
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 errorsThe post does not give a reproducible protocol, test duration, or workload definition, and the results have not been independently verified. They therefore cannot establish how the system would behave at another scale, under a different workload, or in production.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What to learn from this design
PaperPlanes is best read as an example of an architectural direction, not as proof that a particular database or memory scheme solves long-term agent memory. Its design highlights several questions worth answering when building a system that must remember information beyond one interaction:
- Can it represent disagreement? Decide whether contradictory claims are preserved, overwritten, or escalated for review.
- Can it explain what it believed earlier? Historical inspection is valuable when a result needs to be traced to the state available at the time.
- How does old information lose priority? A decay rule needs a stated policy and evaluation; a formula alone does not establish retrieval quality.
- What happens under concurrent writes? Test with a defined workload and report duration, collisions, retries, and lost or rejected writes.
- What does the operational design cost? Persistent, queryable memory adds implementation and maintenance work beyond sending more text in a prompt.
These are evaluation questions, not results from a controlled comparison. Somay’s post does not compare PaperPlanes with large-context prompting, conventional vector retrieval, or alternative databases, and it does not establish cost or retrieval-quality results.
Project status and evidence boundaries
Somay says PaperPlanes was submitted to the CockroachDB × AWS Hackathon. The project description is an author’s account; its code and current state are not independently verified here. The implementation details and test figures should therefore be attributed to the post rather than treated as independently reproduced findings.
Quick Recap
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.
