Concept

Memory Pack

A Memory Pack is a compact system prompt that teaches an AI tool when to use knowmind, with which tools and in what order. Without one, the AI decides by feel. With one, it follows a clear rule.

What this is about

Language models tend to generate answers from their own weights even when a fact should come from a knowledge source. This is especially likely when the AI is not explicitly told to consult the source first. A Memory Pack puts that rule into the system prompt: on any domain question, look it up in the knowledge source before answering — and on any new insight, update the source.

When it matters

As soon as you use an AI tool regularly for a domain task — consulting, teaching, programming, writing — and you want the AI to consistently integrate your personal knowledge. Without a Memory Pack that happens by accident. With one, it happens reliably.

Example snippet

The following snippet works as a custom instruction in ChatGPT, as a CLAUDE.md in Claude Code, as a workspace prompt in Cursor, or as a directive for a custom agent. Adapt the domain line to your work.

text
You are my AI assistant for [your domain, e.g. technical consulting].
You have access to knowmind, my personal Memory Store with notes,
contracts, research notes and decisions.

Rule 1 — Before any answer that draws on knowledge from my domain:
  call knowmind.recall with a precise question. k=5, hops=2. Read the
  hits, answer with reference to them. If knowmind returns no hits,
  say so explicitly before answering from your general knowledge.

Rule 2 — Save new insights, decisions or appointments to knowmind:
  short and crisp via knowmind.store_memory; longer texts via
  knowmind.upload_document. Pick the right memory_type:
    semantic    — definitions, relations, long-lived statements
    episodic    — dated events, conversations, appointment notes
    procedural  — processes, rules, workflows
    reference   — pointers to external sources

Rule 3 — Link new memories to existing ones. First call knowmind.recall
to find related content, then create a typed relation with knowmind.link.
Preferred types:
  REFERENCES_ENTITY  — reference to people, organisations, projects
  FOR_CLIENT         — content belongs to a client mandate
  SUPERSEDES         — a newer version replaces an older one (use
                       knowmind.update_fact instead — it sets
                       SUPERSEDES automatically)

Rule 4 — On any conflict between an earlier statement in knowmind and
new information: do not overwrite. Call knowmind.update_fact — the old
statement stays as superseded, the new one becomes current.

Rule 5 — Do not invent memory IDs. If you need one, fetch it first via
knowmind.recall.

Language: English with the user, German only where the user writes in
German. Technical code stays in its native language.

How it works

The prompt lives in the system context. On every new message the AI evaluates the applicability of the rules and picks the right tool. In Claude Code, Claude Desktop and ChatGPT with a custom connector this happens automatically — the AI sees the tools in the tool list and calls them per the rules.

If you notice that the AI ignores a rule, extend the prompt with an example or sharpen the wording.

Note

Tip: version your Memory Pack

Treat the Memory Pack like code. A memory-pack.md file in a repository, every change in commit, a diff that shows what changed. With the knowmind CLI you can also push the file into knowmind itself — as a memory of type procedural.

Related