Learned Memory

Beyond the admin-authored facts on the Team Memories page, SuperClawd has memory that the AI builds and maintains automatically as your team works. You don't write these by hand — they accumulate from real sessions, and SuperClawd files each one where it belongs so the right context shows up next time.

The memory tiers

SuperClawd keeps memory in a few distinct tiers. One is authored by people; the rest are learned by the AI as it works. Each tier has a different audience.

Tier	Authored by	Who it's shared with
Team	People (maintainers, admins, owners)	Everyone in the workspace
Repo	The AI, captured as it works	Everyone on that repository
Agent	The AI, captured as it works	Anyone who runs that reusable agent
Private	You, or the AI on your behalf	Only you — author-only

Private Memories

Private Memories are per-member and private to you. Only you can see and use them — even the workspace owner or an admin can't read another member's private memories. They're account-scoped: they follow you into every workspace you belong to and outlive any single workspace (they survive workspace deletion), and you manage them on the top-level account Memories page rather than inside a workspace. They're created two ways: you add them yourself in the SuperClawd app, or the AI captures one for you as you work. Either way, they're injected into only your own sessions, so your personal working style follows you without leaking to the rest of the team.

You control your own private memory

Private Memories are account-scoped and follow you across every workspace. You control them with the private-memory toggle in superclawd config (on by default); when you turn it off, no private memories are created or injected for you. There's no workspace-level or owner control over another member's private memory.

Automatic capture

As the AI works, it distills the durable facts worth keeping and routes each one to the right tier on its own:

A build command or a repo gotcha becomes a Repo memory, shared with everyone on that repository.
A personal working preference becomes a Private memory that only you ever see.
A lesson tied to a reusable agent's role stays with that Agent.

You're never asked where to save anything — the filing happens in the background while you keep working.

Judged on your own Claude — zero extra inference

SuperClawd does no LLM inference of its own to decide what to capture. The judging runs on the same Claude you're already using — same privacy stance as the rest of the product: content-free, team-owned, and fully inspectable and editable in the app.

How learned memory reaches your session

SuperClawd doesn't pour every learned memory into your session. When a session starts, it loads a small, always-on CORE — your most relevant lessons, in full — alongside a compact INDEX of everything else: just the titles. The AI pulls the full text of an indexed memory only when it actually needs it. That keeps your session focused on what matters right now instead of carrying every memory at once.

Found by meaning, not just words

Recall now matches on meaning, not only on exact wording. A search for "expired auth" will surface a lesson written about "sessions rejecting past-expiry tokens" — even though they share no keywords — because SuperClawd combines semantic similarity with classic keyword matching and fuses the two. You get the right memory whether you remembered the exact phrasing or just the idea.

Surfaced automatically

Relevant memories show up on their own — you don't have to ask for them. When you send a new prompt, SuperClawd looks for memories that match what you're asking, and when a file is opened it surfaces memories about that file. The matching context appears in the session automatically, so the AI applies what it already knows without you prompting it.

You can turn it off

Automatic surfacing is on by default. If you'd rather memories only load when explicitly searched, set SUPERCLAWD_AUTO_RETRIEVE=0 in your environment.

Memory decay

Learned memories shouldn't pile up forever. An AI-learned memory you stop using is quietly retired after about 90 days — and retiring is fully reversible. The clock resets the moment a memory is used again, or as soon as you lock it. Team Memories (human-authored) and any locked memory never decay. In the app, an active memory shows a "Decays in N days" countdown, and a memory that's been retired for going unused is marked "Retired (unused)" — you can always bring it back.

Reviewing what's been learned

Nothing the AI captures is hidden. Every learned memory — Repo, Agent, and your own Private Memories — is listed in the SuperClawd app, where you can read it, edit it, or delete it. If the AI files something you'd rather it hadn't kept, remove it and it's gone.

Billing

Pulling the full text of an indexed memory uses get_memories. On paid accounts, a get_memories call that returns one or more repo or agent memories is metered at a flat 0.10 credits per call (free on the free tier). Private memory reads are always free (account-scoped and unbillable), and team memories are free because they're injected at session start rather than fetched. Keeping memories is never charged. See Billing & Credits for details.