AI Agent Memory Files Rarely Read: Only 5 of 386 Sessions Opened Lessons File

·4 views

An AI agent's logs showed its lessons file was opened in just 5 of 386 sessions over several months, meaning notes written at session end were seldom read when needed. A separate recall step surfaced relevant memory files on 26 of 370 prompts, yet none of the pointed-to files were subsequently opened. Another agent found its memory was injected into the system prompt at session start rather than opened by tool calls, so standard open-counting missed it entirely. Her injection tests found the content was present in every sampled request but showed apparent use in only a minority of cases, with caveats about the sampling method. The authors recommend pairing each memory store with an explicit read trigger and a way to detect when reads stop.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Cursor-based pagination for multi-tenant RAG ingestion at 10k documents

The article argues that ingestion for a multi-tenant help center should use tenant-scoped cursor pagination over a stable snapshot rather than offset pagination, which becomes a correctness and index-cost problem as content grows. The author describes ingesting documents in bounded pages ordered by update time and document ID, then chunking and upserting vectors with deterministic keys so repeated runs stay idempotent. Page size and chunk size are presented as tunable decisions to be measured against the document mix, not fixed constants. Metadata such as tenant, revision, and access controls is kept with each chunk so retrieval can enforce tenant boundaries before ranking. The piece is written from the perspective of a logistics SaaS workload with carrier rules, customs notes, and warehouse procedures.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Go Review Jobs: Handling Long Document Extraction and Timeouts Without False Completion

The article explains how to run bounded review jobs that turn long documents into JSON while keeping an explicit incomplete state after a timeout. It argues that a parseable response does not prove a review finished, so deadlines, token budgets, and evidence coverage should be tracked separately. Failures such as local deadlines, transport errors, context-budget rejections, and validation problems are recorded in distinct categories because each calls for a different fix. The author advises against blind retries, since a timed-out request may still have run remotely, and recommends deduplicating published findings by review revision and finding identity. Chunking should follow source boundaries and fit the selected provider's tokenizer limits, with file paths and line numbers kept outside the model's output.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Invisible Unicode Characters Can Quietly Break Email Templates and Filters

A scenario is described in which marketing copy pasted into an email template passed a visual check but caused a rendering problem and a missed filter match. Invisible Unicode characters such as direction controls and zero-width characters occupy positions in text without displaying anything, and they often arrive unnoticed through copy-paste from documents or web pages. These characters can reorder displayed text, as in the file-name disguise example, and can split keywords so content filters fail to match them. The article notes that a 2021 research effort, known as Trojan Source, showed similar tricks could make source code appear different to reviewers than to compilers. Developers can detect such characters with a regular expression in CI pipelines, though the article cautions that some invisible characters are legitimate in certain languages and should not be blindly removed.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

CVE-2026-79875: Critical Auth Bypass in HPE ArubaOS AOS-8 PAPI Protocol

A critical authentication bypass, tracked as CVE-2026-79875, has been identified in the PAPI protocol of HPE Networking ArubaOS AOS-8. The flaw can be exploited by an unauthenticated attacker on an adjacent network, potentially granting administrative access, enabling configuration changes, and exposing or altering sensitive data. Affected versions span 8.13.0.0 to 8.13.3.0 and 8.10.0.0 to 8.10.0.23, with fixes available in 8.13.3.1 and 8.10.0.24. The CVE record gives it a CVSS 3.1 score of 9.3 and classifies it as improper authentication (CWE-287). Asset owners are advised to inventory device builds, restrict reachability of the management interface, apply fixed releases or document compensating controls, and review the wider HPESBNW05162 advisory and audit logs for unauthorized administrative activity.

0 comments Read more at DEV Community