01
You approve what it learns
New capabilities never switch themselves on. They arrive as proposals with a description you can read, and become active only when you approve them — rejecting one takes a single command.
Governed memory · by Consultancy in Action
An AI assistant that learns from you — safely.
Out of the box, an assistant forgets everything between sessions. Self-Improve gives it a governed memory: it reviews each session, proposes what to remember, and only learns what you approve — every write journaled, one file pausing the whole loop. Local-first, £0 to run.
- Consent-first
- Fully journaled
- One-file kill switch
- £0 to run
A governed memory & learning loop for AI assistants, at a glance
Consent-first
You approve what it learns
New skills arrive as proposals with a description you can read. Nothing self-activates.
£0 to run
On the subscription you have
Plain files, no database, no API keys, no cloud service. Local-first by construction.
Any agent
One loop, swap the storage
Claude Code, AEGIS, Codex, Hermes, or a local model — the methodology is agent-neutral.
Why it exists
A learning assistant is a governance problem before it's a feature.
Out of the box, an assistant forgets everything between sessions. You repeat the same corrections, restate the same preferences, re-teach the same techniques — and every repeated correction is paid for twice: once in time, once in trust.
The obvious fix — let the assistant rewrite its own instructions — is also the obvious governance nightmare. Unsupervised self-modification is exactly the failure mode responsible AI operations exist to prevent.
The real problem isn't making an assistant learn — models do that eagerly. It's making one learn safely: never the wrong lesson, never a silent rewrite, and always under a human who can see it, veto it, and stop it.
How the loop works
Review → propose → you decide → curate.
Two kinds of learning, kept separate: memory (who you are, the state of the world) and skills (how to do a class of task). Every step is journaled.
01
Review
At the end of a session — or on demand — the loop reads the conversation against a written methodology, looking for three signals: you corrected the assistant, a non-trivial technique emerged, or an existing instruction proved wrong.
02
Propose
Findings become drafts. Corrections and preferences become memory (who you are, the state of the world); reusable techniques become skills (how to do a class of task). Named for the class of task, never today's specific bug — and nothing goes live.
03
You decide
In consent-first mode a new skill is a proposal you approve or reject with a single command. Approve promotes it into the live library; reject archives it, dated. One file pauses the entire loop instantly.
04
Curate
A weekly pass merges overlapping learned skills and archives stale ones, so the library stays sharp instead of silting up. Its AI pass has no shell — it only proposes moves a validating script checks before applying. Every move is journaled.
The full command surface — review, approve, reject,
curate, pause — lives on the commands page →
The governance is the product
Five safeguards, in plain English.
"AI that improves itself" is usually a demo or a liability. This is the third thing — a learning system a business could actually operate, because every way it could go wrong is gated.
02
One file stops everything
A pause flag kills every automatic review and curation instantly, no uninstall required. Manual use keeps working; the automation simply stands down until you say otherwise.
03
The reviewer runs in a straitjacket
The background reviewer gets read-only tools plus write access to exactly two areas — memory and proposals — under a 15-minute hard cap. It cannot edit your files, settings, or any skill: every skill change is a proposal you approve. That's enforced by its tool permissions, not just policy.
04
It can't touch what it didn't create
Every learned skill carries provenance in the file itself. Anything without that stamp — your own skills, bundled ones, a protected list — is invisible to the loop and the curator alike.
05
Nothing is ever deleted
Archive is the maximum destructive action: rejected proposals and retired skills are moved, dated and kept. The journal records every action — including the sessions where it chose to save nothing.
By the numbers
What was delivered — verified facts, not projected returns.
3
build phases shipped
manual review → gated auto-review hooks → weekly curator
7
commands in the loop
review · status · approve · reject · curate · log · pause/resume
2 areas
reviewer's write scope
headless, read-only tools; writes limited to memory + proposals under a 15-min cap — skill changes are proposals you approve, never live edits
100%
of writes journaled
append-only journal; archive-only — nothing is ever deleted
£0
marginal cost to run
plain files + the subscription you already pay for — no DB, no keys, no cloud
4+
agents it runs on
Claude Code, AEGIS, Codex, Hermes, or a local model — one methodology, swap the storage
We'll be honest about its age: on our own Claude Code deployment the journal is a handful of entries deep, one learned skill is live so far, and the automation is currently paused — by choice, with the one-file kill switch. That's the control working exactly as designed.
How it was built
Three phases, three days — automate the trigger, never the trust.
Phase 1
The control plane, deliberately manual
A single command reviews the session against a written methodology and writes memory + skill proposals. Updating-in-place beats duplicating; a do-not-capture list keeps transient failures out; in consent-first mode a new skill never goes live. Every action — including “nothing to save” — is appended to a tamper-evident audit journal.
Phase 2
Automate the trigger, not the trust
Session-end hooks call a gatekeeper that only fires a review when thresholds are met — enough user turns, tool calls, or repeated correction signals — under a daily cap, a per-session gap and a single-review lock. The reviewer sees a redacted digest and runs headless with tools cut to read-only plus two writable folders, under a 15-minute cap.
Phase 3
Keep the library from silting up
A weekly curator merges overlapping skills and archives stale ones. Its AI pass has no shell at all — it only proposes moves in a plan file; a validating script checks provenance, pins and protection rules before applying anything, then commits the library to git. Nothing is ever deleted; skills the loop didn't author are untouchable.
“AI that improves itself” is usually either a demo or a liability. This is the third thing — a learning system a business could actually operate, auditable end to end, running at £0 marginal cost on the subscription already paid for.
| Route | Typical cost | Time to ship |
|---|---|---|
| Specialist AI-platform consultancy | £40k–£80k | 2–4 months |
| Consultancy in Action — AI-accelerated | A fraction · £0 to run | 3 days, 3 phases |
The command surface
Run it interactively or headless
Type /self-improve mid-session, or drive it from a script with learn review / approve / curate / pause. Consent-first, journaled, pausable.
See the commands →
Take it to any agent
One methodology, any assistant
The loop is agent-neutral markdown — only the storage paths differ. Adapter table for Claude Code, AEGIS, Codex and more, plus a one-paste pocket version.
How it travels →
Private preview
Give your assistant a memory you can govern.
Self-Improve is in private preview while we harden it on our own work. Leave an email for early access — or talk to Consultancy in Action about wiring governed memory into your assistant or agent fleet.