Skip to content

Inside-Out Memory — 5-minute reproducible demo (synthetic)

If you want the stricter trust-gating proof first, start with trust-aware context pack proof.

This demo is synthetic (no private/user data). It is the shortest clean showcase for the openclaw-mem wedge:

  • a small set of stable preferences / constraints is stored once
  • relevant memory can be packed into a compact, cited bundle on demand
  • an agent can answer consistently without bloating chat context

In product language, this demo proves a simple claim:

openclaw-mem helps the agent recover what still matters, with receipts.

Who this demo is for

Use this demo when you want to show any of the following fast: - why local memory is more trustworthy than a vague "AI memory" story - how a compact recall pack can beat dumping giant memory files into context - how to keep stable constraints visible without turning memory into a black box

Prereqs

  • uv
  • In this repo: uv sync (or let uv run ... build on demand)

Run

# From repo root
./scripts/inside_out_demo.sh

You should see a packed bundle (plus an optional trace) for a keyword-style query that works even with no API key / no embeddings:

timezone privacy demo style

Demo talk track

Use this 3-beat flow when showing it to someone:

1) Before memory

Ask for a demo plan with no recalled constraints.

The typical agent answer is generic. It may ignore timezone, privacy expectations, or preferred output structure.

2) Pull the recall pack

Run the demo and show the packed bundle.

The point is not raw retrieval volume. The point is that a small, relevant, cited pack comes back instead of a giant wall of memory.

3) After memory

Ask for the same plan again, now with the recalled constraints in view.

The agent should: - use the preferred timezone for time references - keep the demo synthetic / privacy-safe - structure output in the preferred style

That is the core value: not "more memory", but memory that changes the answer in a controlled, inspectable way.

What to look for

  • Timezone preference is recalled (UTC+8 / Asia-Taipei in this synthetic example).
  • Privacy constraint is recalled (demo uses synthetic data; do not leak private notes).
  • Style preference is recalled (index-first / bounded reveal).
  • Pack stays compact instead of dragging in unrelated memory.

Before/After (illustrative transcript)

Before (no memory)

User: “Write a demo plan for my agent system.”

Agent (typical): gives a generic plan, may ignore timezone + privacy constraints.

After (with packed memories)

User: “Write a demo plan for my agent system.”

Agent (memory-aware): - uses the preferred timezone for time references - explicitly keeps the demo synthetic / privacy-safe - structures output index-first and cites the recalled constraints

Why this demo matters

This demo is intentionally small.

That is a feature, not a weakness: - it is a vertical slice of the recall contract - it is reproducible - it is safe to run publicly - it gives a clean before/after story for README, docs, and live demos

If you want a first GTM-friendly demo for openclaw-mem, start here.

Architecture diagram

  • Mermaid: docs/showcase/inside-out-architecture.mmd