Memory Base
A reusable memory infrastructure that keeps what should be believed, updated and recalled over time.
Existing memory systems may already be enough for many people.
I want a layer whose storage, belief updating and recall I can control and reuse — in agents, tutors or apps. Claude Code is the first host; Codex reads from it.
System sketch — not the final architecture.
- Exists now
- Personal and project information persists across sessions and is recalled later.
- current_memory updates when understanding changes; append-only events keep the trace behind it.
- M1 (core mechanics) and M2A (safe control plane) are closed.
- Still unproven
- Whether compact understanding adds value beyond episodic recall.
- Fewer repeated mistakes, or any advantage over Claude Memory, claude-mem or project memory.
- Procedural memory — remembering ways of working. Still an idea.
A debugging episode
Recall quality dropped badly. The suspect was a 25-row cap — it was not the cause. Values written by the producer ran about 221 characters, against about 106 for ones people wrote, so only about five rows fit in the recall budget.
Since then, a rule that matters gets a hook or a gate, not only prompt text.
What turned out wrong
- A token / compression advantage, as first assumed.
- Path or symbol proximity as relevance.
- That existing systems cannot carry context across sessions.
Persistence, updating and recall work in daily use with Claude Code.
Experiment phase: M1 and M2A closed; the comparison against existing memory tools is still being tested.