Blog
Technical deep dives, architecture decisions, and lessons learned building Comis.
The memory that models you - what we shipped, measured at $0
Comis now ships a per-user profile, a per-channel relationship model, an opt-in tool that answers a question from memory grounded in cited recalled records, a loop that learns which memories prove useful and bounded-tunes recall ranking, and principled ranking decay of stale memories - each proven keyless at $0. The one measured learning signal is a +0.1 recall-score lift over 5 episodes (rank flat); the re-stated cross-judged baseline stays separate, and the costed competitor comparison is honestly deferred.
The honest agent-memory benchmark: we built the proving machine first
Agent-memory leaderboards are easy to game, and the judge model alone can swing a score from ~49% to ~94%. So before publishing a number, we built the machine that proves one: an open, reproducible, cross-judged harness. Here is what we measured at $0, what we deliberately did not publish, and the one command to reproduce the full head-to-head yourself.
How One Sentence Spawned Six Sub-Agents and a Trade Idea
I typed one sentence into Comis. It built a 6-node DAG, dispatched six isolated sub-agents, fired 101 tool calls, ran 64 web searches, and came back with a trade plan that included three entry tranches, a stop, and a 4:1 reward-to-risk. Here's how the pipeline orchestration actually works.
How I Cut Our Multi-Agent Pipeline Cost by 64% with Cache Forensics
The story of debugging Anthropic's prompt cache behavior in production, discovering three root causes, building a cache fence system, and uncovering a TTL monotonicity violation - the journey from that first 64% win to today's 81% cost reduction.
Why AI Agent Security Is Hard (And What We Built to Solve It)
Most AI frameworks bolt security on after launch. Comis was designed around a different question: what happens when an AI agent has real power and someone tries to abuse it? Layered runtime defenses later, here's what I learned.
Anatomy of a Context Engine: 8 Layers That Keep AI Agents Affordable
How Comis strips thinking traces, evicts dead content, masks old observations, compresses long conversations, and keeps useful detail recoverable without turning every request into a transcript dump.