- Author
- Tim Kellogg
- Published
- 15 June 2025
- Kind
- Blog post, largely AI-written by the author's account
- Original
- timkellogg.me/blog/2025/06/15/compression
- Subject
- File-based agent memory
- Themes
- What lasts
What it is
Tim Kellogg's post compares three ways of working within limited context. They are Letta's layered memory, Anthropic's multi-agent research system and Cognition's single-agent advice. Kellogg says OpenAI Deep Research formatted a conversation into the post. Kellogg also reports light editing, and calls the result generally all AI.
What it says
- Letta splits memory into core, message buffer, archival and recall blocks.
- The post describes each block as a lossy compression of the last, like CPU caches.
- Anthropic's system compresses in parallel, with subagents reporting to a lead agent.
- Cognition's Walden Yan says agents share context poorly. Yan suggests one agent plus a model that condenses its history.
- The post closes by treating compression as a form of cognition.
What evidence it gives
The post offers no measurements of its own. It relays Anthropic's report that a multi-agent setup did better on broad queries. That report and Yan's argument come from linked posts.
Interests and limits
Letta's archive is often a vector database, so this is not a file-based design. It appears here as a frame for deciding what a memory layer keeps. The text is largely AI-written, by the author's account.
Connections
These links are this wiki's own reading, not claims the sources make.
- Visualizing 100,000 Years of Earth in WebGL. Kellogg treats memory as a budget, where detail is spent on what matters most. Stam does that literally, giving 206 of 256 height values to the coastline's range. So what is a memory file's shoreline?
- Just-In-Time Agent Memory with Runtime Agentic Research. Kellogg treats memory as layers of lossy compression. Yan and colleagues say compressing early discards details that later matter. They keep raw sessions and search them. The choice is about when to pay: when writing, or when asking.
- Context engineering. Kellogg reads multi-agent research as an exercise in compression. Loy says agents hand off token sequences, and calls that hand-off an API contract. Then a contract is a statement of what a compression must preserve.
- The two Christian saints who are secretly the Buddha. The Buddha's story reached Christian saints through many retellings, and Bodhisattva became Budhasaf. Kellogg treats memory layers as lossy compression. The legend shows what can survive many hand-offs: plot, not provenance. Historians took centuries to accept the source.