Building memory for agents, in public.
Design decisions, benchmarks, and the flows that make the loop feel instant.
Importance, not freshness: why we reward stable code
Recency decay alone starves agents of the core abstractions they need most. Here's the stability bonus, and why it ramps over 30 days.
Read post →How we hit 12–15× token compression without losing accuracy
Compression levels L0 → L3, greedy budget packing, and the 0.15 swap threshold that keeps context stable across turns.
Read post →The edit fence: draining agent edits in under 50ms
Why we replaced a 300ms Redis polling loop with an in-memory sync.Map — and how it removed the last stutter from the agent loop.
Read post →Why we chose tree-sitter over LSP
2–3× faster than WASM, incremental, error-tolerant, and it works on files that don't compile — which is most of the files an agent touches.
Read post →Agents are untrusted writers: Git-style branches for memory
Shared base pool plus per-agent branches. Soft scoping, semantic conflict detection, and four merge strategies.
Read post →Tokens are currency
Every byte sent to an LLM has a cost. Optimize for information density, not completeness — a five-principle manifesto.
Read post →