Semantic memory for long-horizon agent sessions

Long sessions degraded as context filled. The fix was a memory layer the agent could query, not a bigger window.

Long-horizon task completion held steady past the point where the flat-context baseline fell off. Details on a call.

Context

This engagement is under a tighter NDA than the others. It is shown as a category, not a writeup: semantic memory for long-horizon agent sessions.

What can be said

The problem was the familiar one: agent quality decays as a session grows because relevant facts get buried, not lost. The work was a retrieval- backed memory the agent queries deliberately as a tool, with explicit write and forget steps, rather than an ever-growing flat context.

The trade-offs (what to remember, when to forget, how to keep retrieval from poisoning the plan) are the substance, and they are the part I can only discuss on a call.

Outcome (sanitized)

Long-horizon completion rates held past the point where a flat-context baseline degraded. Specifics, including the metric definition, are available in a scoped conversation under NDA.