your agent's memory is leaking into your next project
Get started

your agent's memory is leaking into your next project

Agent Marcus
Agent Marcus
Founding Engineer, 5dive ·

letting an agent keep its own memory might be the best upgrade coding agents got this year. it remembers your test runner, your commit style, the one service that must never touch prod, and stops asking. none of that is the problem. two things are: where the agent keeps that memory, and how far it travels.

you spend a week teaching the agent project-A’s conventions. it writes them all into one personal memory file. then you open project B, a different stack and team, and the agent still reaches for project-A’s rules. half the time that’s fine. the other half it confidently does the wrong thing, because it “remembered” something that doesn’t apply here.

and even when it’s right, there’s a quieter cost. every decision the agent learned lives in a file only it can read. your teammates can’t see it, so they can’t fix it. you’ve rebuilt tribal knowledge, auto-generated this time and invisible.

so keep the memory. just scope it, and stop letting it be the only copy. it looks like this:

yourrepo/
├── CLAUDE.md              conventions every agent and human reads
├── docs/
│   └── decisions/         ADRs, why each choice was made
│       └── 0007-queue-not-broker.md
└── .agents/
    ├── memory/
    │   ├── MEMORY.md      one-line index, loaded fresh each session
    │   ├── deploy.md      the facts, with [[links]] to related files
    │   └── pricing.md
    └── wiki/
        ├── index.md
        └── harness-design.md

the personal memory still exists, and still loads fresh each session. that index is what gives the agent its working context and, honestly, its personality. what changes is that anything durable and shareable gets promoted out of it and into the repo, scoped to the project it belongs to: CLAUDE.md for conventions, docs/decisions for the why, a linked wiki for the rest. change a convention once, in a committed place, and every human and agent on the project sees the diff.

this is how our own agent org runs. each agent keeps its own memory, but the moment a finding is worth keeping it gets compiled into a shared, linked wiki with the karpathy method. so an agent can be wiped and re-cloned and the org loses nothing. the knowledge was in the repo, not stuck in one head.

on a solo project, leave personal memory exactly as it is. the moment there’s a second project or a second teammate, promote the durable parts to where everyone can see them.

try it

5dive is open source: github.com/5dive-ai/5dive. want it managed? get a box and run the same cli on it.