Memory is advisory, never authority.
It retrieves and interprets history, with citations. It can never approve, spend, sign, or publish anything.
Projects / Agent reasoning, memory, and governance
EXPERIMENTAL
A memory that outlives the session, and a mind that tidies it.
e3d-brain
A local-first MCP memory server over a plain folder of Markdown notes, so Claude and agents keep continuity between sessions. A second layer looks after the memory: it notices what was superseded, duplicated, or stale and proposes changes for a person to accept. Memory is advisory, never authority.
Both layers are built and tested (664 tests). It has not had a human security audit, has only been verified on macOS, is not yet used day to day, and the integrations with the other projects are planned, not live.
git clone https://github.com/spacepacket1/e3d-brain.git
▶ Watch the concept film (1:52)Every AI session starts blank. Claude Code forgets last week's decision. An agent repeats Tuesday's mistake. A team re-explains what it already settled. The records exist, but the people and agents doing the work cannot reach them at the moment they matter.
Two ideas hold it together
Models propose; humans decide.
Nothing becomes permanent memory, and no existing note changes, until a person says yes. Every change is undoable.
Memory, then cognition
15 phases built
Memory layer
A small, local-first MCP server over a plain folder of Markdown (an Obsidian vault). Ranked search over a rebuildable SQLite index, with every result carrying a stable id, a revision, and a citation pinned to that exact revision so stale citations can be detected. Three fixed modes: inbox for a person's Claude, journal for an agent's own expiring log, and read-only for a curated brain.
9 phases built
Cognition layer
Owner-run jobs that read the brain and write typed, cited, previewable proposals: which decision superseded which, what is duplicated, what real outcomes taught, what has gone stale, and what keeps being asked with no good answer. Nothing is applied by a job, and no MCP tool can start one.
ebrain review list # typed proposals, with authentication and taint
ebrain review diff <id> # the exact change against the current notes
ebrain review accept <id> # revision-checked, crash-safe, undoable
ebrain review reject <id> # remembered, so it is not proposed again
ebrain review undo <id> # restores what accept changedWhere it fits
Memory is not the record. It is the reader of the record.
e3d-corpBusiness events: proposals, approvals, spend, outcomesAuthority for: The business record
e3d-pilotThe idea ledger: specs, debates, PRs, verified resultsAuthority for: The engineering record
e3d-brainNotes that point back to bothAuthority for: Nothing. It helps retrieve and interpret.
What is not done
- No security review. The code was written by AI and checked by tests and independent model reviews of the design, not a human security audit.
- Not yet connected to real clients or the rest of the ecosystem. The tests drive the real MCP server over stdio, but integrations are planned, not live.
- Not yet proven useful. It is built to measure whether the cognition jobs earn their keep; it has not run long enough on real data to say.
- Verified on macOS (Apple silicon). Not yet run on Linux.
Works with
Want this applied to your own workflows?
We build these to help others. If you want a system like this working in your business, that is what our consulting is for.