This started as a simple question: how do the tools we use every day actually remember? So we pulled apart Codex's memory engine and Claude Code's, set them side by side, and went looking for insights we could actually use — specifically, to improve the cross-agentic memory inside OpenBrain. Sharing memory across agents is one of the reasons OpenBrain exists, and what these two systems get right (and wrong) is a real opening to make it better. This is an improvement to that — or it could be.
There's a second thread, and it doesn't necessarily touch OpenBrain at all: a personal agent fleet — a project that means a lot to me and is already pushing some genuinely cool stuff. It's had agents come and go (a video producer for a while, a few others), it gets real work done day to day, and giving it a smarter shared memory would be a real upgrade. Two places to apply it; one set of lessons.
The thread that runs through all six pieces: agents propose memory, and a consolidation layer owns what becomes a durable fact. Almost everything good follows from that one boundary.
The synthesis the rest of the series builds toward: a single memory layer that doesn't belong to any one model. Host it in Supabase, put an MCP interface in front, and every agent — Claude, Codex, Gemini, Grok — reads and writes the same brain, with provenance baked in. It's the shape both applications below borrow from, and the clearest statement of the propose-then-own principle.
Read the vision →Two real systems, torn down — because you can't improve cross-agent memory without first understanding how the best tools already do it.
A teardown of a live Codex memory export. It's two tables and a promotion gate: each session is distilled into a structured record, only the selected ones get promoted into durable memory, and a leased, retryable job ledger keeps the whole pipeline crash-safe. Memory treated as a process, not a pile — this is where most of the borrowable ideas come from.
Read the teardown →The opposite philosophy: files, not a pipeline. A hierarchy of human-readable CLAUDE.md instructions loaded at the start of every session, plus a model-managed directory of fact files recalled by relevance. Everything is plain text you can read, edit, and commit to git. Strong where Codex is weak — legibility and version control — and weak where Codex is strong.
Read the teardown →The two approaches head to head — pipeline-and-ledger against file-and-hierarchy — across ten dimensions, with the clear winner marked on each. The point isn't to crown one; it's to see exactly which strengths to steal from each, and to notice the gap they share: neither is portable across tools. That gap is the whole opportunity.
Read the comparison →Same lessons, two independent homes — one tied to OpenBrain, one not.
OpenBrain already aims at cross-agentic memory — that's part of what it's for — so these are improvements to something real, not a new product. Three patterns transfer almost for free: a promotion gate so captures become curated knowledge instead of a dump, a job ledger so its ingest and wiki pipelines are crash-safe and resumable, and provenance so it can evolve a fact without quietly contradicting itself. The aim is to make the cross-agent memory OpenBrain is already reaching for genuinely better.
Read the OpenBrain ideas →The other thread, separate from OpenBrain: the personal agent fleet already has most of a good memory system — a knowledge graph, a provenance table, a working roster of agents — but it's fragmented, and it's missing two pieces. v2 keeps everything that works and adds just two things: the propose-then-own boundary, and a real job ledger, rolled out in low-risk phases. A concrete, modest upgrade to a project that's already accomplishing a lot.
Read the fleet proposal →Studying how Codex and Claude Code remember was never the point — it was the means. The point is a smarter cross-agentic memory: better inside OpenBrain, where it's already part of the mission, and better in a personal fleet that's quietly been doing real work for a while. Start anywhere above; the vision ties it all together.