Agent memory handoffs cut replaced context 98.65% by moving verbatim text out of the prompt

The per-run averages make the headline figure concrete. Each replaced window held about 111,062 characters, and the handoff that replaced it about 1,501 ^[1]. According to the author of the audit, the 98.65% comes from summed lengths across all runs, not from an average of per-run percentages, so the longest windows count most ^[1][2].

Here is what happens to a window before any of that saving shows up, per the post:

1. The agent selects a message-safe middle window and cuts its messages into role-aware segments of at most 3,800 characters ^[8]. 2. Tool, system and developer segments are protected, along with recognized patterns such as code fences, commands, paths, hashes and error evidence. These resolve to KEEP before any model classifies them ^[9]. 3. Everything else goes to a GLiClass router. Weak decisions and model failures fall back to KEEP ^[10]. 4. COMPRESS candidates pass through LLMLingua and then an NLI gate ^[10].

Every default in that sequence leans toward keeping text. Compression ended up handling about 0.003% of segments ^[2]. A stage that fires that rarely is, for accounting purposes, a KEEP stage with extra dependencies. The context saving comes from the handoff itself: once a window's retained material is durably stored, the whole window leaves the active context, KEEP segments included ^[6]. "The large active-context change comes from externalization," the author wrote ^[7].

I think falling back to KEEP is the right default for agent work. A misrouted classifier costs storage. A misrouted compressor can cost the exact port a later tool call needs. The planner smoke test in the deployed API environment exercised that case: 24 synthetic tool messages, each with an exact endpoint, port and command, were all retained verbatim and still yielded a much smaller handoff ^[17]. That test made no database writes, and its timing was a single observation ^[17].

The hash audit establishes that the stored text is intact. Whether an agent can use it has been tested once, on synthetic data with a live answering model and real embeddings ^[11]. Set aside the question whose correct answer was UNKNOWN, and the handoff-only arm scored zero of 11 ^[3]. The recall arm matched full history while using substantially fewer input tokens, retrieved evidence included ^[12]. "This is a small controlled result, not a claim of general lossless agent memory," the author wrote ^[14].

Repeated runs can contain related material, so the 1.14 billion source characters are not a count of unique information ^[1][16]. For the 98.65% to carry over to another deployment, its windows would need a similar size profile, and its agent would need to decide for itself when to recall. In the synthetic test that decision belonged to the harness, which made the retrieval calls ^[13].