A 404 from a memory API can mean two very different things: the client is calling the wrong route, or the app is asking about a contact whose memory store has never been created. Antony Sebastian’s account of building Promise-Keeper, published on DEV Community on September 29, 2026, separates these cases clearly, and the same separation is the most useful habit to take from it. The app is a Streamlit application that uses Hindsight for memory and Gemini to extract promises and prepare meeting briefs. The author calls Hindsight’s REST API directly, without an SDK.
Start with the errors, because they point to different problems
The article describes three groups of failures: 404 responses from the memory API, timing problems between saving and recalling memories, and failures in the LLM provider layer. Each needs a different response, and treating them as one category is the mistake the account is most useful for avoiding.
As an Amazon Associate I earn from qualifying purchases.
| Symptom | Cause reported in the article | Response in Promise-Keeper |
|---|---|---|
| 404 on a retain call | The author’s first guessed route was wrong. The route the app eventually used was documented in the article as /v1/default/banks/{bank_id}/memories. |
Corrected the path and request body to match the API’s documentation. |
| 404 on recall for a new contact | The contact’s memory bank did not exist yet, which is a known first-use condition in this app. | Treated as an empty list, so the first meeting brief starts empty. |
| Recall missing a fresh save | Hindsight processes retained text with an LLM, so a save can take several seconds to become available. | Set generous request timeouts and ran brief generation only after the save completed. |
| 403 responses from Groq | Requests to the Groq provider were blocked. | Kept the provider and model settings in an environment file so they could change without editing application code. |
| 503 high-demand responses | A provider-side capacity incident. | Retried selected 5xx responses with increasing waits. |
Route and body: check the documentation first
The author first guessed a REST route from common conventions and received 404 responses. The lesson drawn is simple: read the service’s current documentation for endpoint paths and request bodies before writing the client. The retain call in this app used the following shape:
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallPOST /v1/default/banks/{bank_id}/memories
{"items": [{"content": "..."}]}This is the route and body the article reports for its implementation. It is not a substitute for the official Hindsight reference, and the article does not confirm that these values remain current.
#1 Best Overall
The new-contact 404 is an empty state, not a failure
Promise-Keeper creates one memory bank per contact, with names such as contact_priya_sharma. When a user asks for a brief about a contact who has never had a meeting, the bank does not exist yet, and recall returns a 404. The app handles that condition as a normal first-use case and returns an empty list.
The author’s summary of this behaviour is a useful line to quote: “A 404 isn’t always an error.” It describes this app’s first-recall path. It is not a general HTTP rule. The article does not establish that a 404 from Hindsight always means “no memories”, and it does not suggest suppressing every 404. An app that treats all missing resources as empty will hide real routing mistakes, so the exception should be limited to the specific case the app knows about.
Rank #2
- Used Book in Good Condition
How promises are stored and reconciled
A promise is saved as a plain sentence containing the date, the recipient, the task, the due date, and an open status. When the promise is fulfilled, the app adds a separate fulfillment memory rather than editing the original. A recall query returns both records, and Gemini reconciles them into a current status.
This append-and-reconcile pattern works for the author’s app, where the sample recall question is “What promises are open or overdue?” The article does not present it as a general best practice for memory systems. Editing in place, or storing structured status fields, would be reasonable alternatives that the account does not evaluate.
Timing: do not assume a read follows a write
The article reports that retain can take several seconds, because the text is processed by an LLM before it is indexed. An immediate recall could occasionally miss a fresh save. The app therefore does not assume a synchronous read-after-write. It sets generous request timeouts and makes save completion happen before brief generation starts.
These are the author’s observations from this project. The article does not describe a documented service-level guarantee for indexing time, so an application that depends on tighter timing should measure its own behaviour and check the official documentation.
Provider failures and retry behaviour
The LLM provider layer caused several problems in the same project. The article reports Groq requests blocked with 403 responses, a Gemini model that became unavailable to new users, and a 503 high-demand incident.
The author’s responses had three parts:
- Selective retries. Certain 5xx responses were retried with increasing waits. The article describes this as one implementation, not a universal retry policy.
- Configurable providers. Provider and model settings lived in an environment file, so a model could be replaced without a code change.
- Fewer calls. Extraction and fulfillment checking were combined into one model call to reduce request use.
The author also reports that the free tier in use was capped at 20 requests per day. This is a dated, author-reported figure from 2026, not a verified current quota. Current Gemini limits should be checked against the provider’s own pricing and rate-limit pages before planning capacity.
Best Value
Local setup problems that look like API problems
Some of the debugging time went to the development environment rather than the API. The article names three causes:
- Windows PowerShell execution policy blocking virtual environment activation.
- A
.envfile accidentally saved with a.txtextension, so its settings were not loaded. - Running a different
app.pyfrom the one that had just been edited.
When a 404 or provider error does not match the code you just changed, check these local conditions before rewriting the client.
Checklist for building on a memory API
- Confirm endpoint paths and request bodies in the service’s current documentation before coding against them.
- Decide what a missing resource means in your application. In this case, a missing bank on a first recall was an empty state, and every other 404 was still treated as an error.
- Assume saves are not instantly searchable. Sequence brief generation after the save completes, and use timeouts that allow for LLM processing.
- Retry only the errors that are likely to clear, and keep provider and model choices in configuration.
- Check the local environment (activation scripts, file extensions, the file you are running) when behaviour contradicts the code.
What this account does and does not establish
The article is a first-person case study of one Streamlit app. It does not compare competing memory APIs or SDKs, and it does not measure how often these failures occur or how reliable the service is. It also does not verify current Hindsight routes, error semantics, indexing guarantees, or provider quotas. Use it to generate the right questions for your own integration, then answer them from the official documentation and your own tests.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Source: Antony Sebastian, “What 404s Taught Us About Building on a Memory API,” DEV Community, September 29, 2026.
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.
