Trust-aware context pack copy pack¶

Use this file as the durable outward-facing copy source for openclaw-mem.

Scope note: maintainer-facing copy source in a public repo; this file is used to shape published copy but is not itself a user tutorial page.

Product positioning¶

openclaw-mem is not trying to remember everything. It is a trust-aware context packing layer for OpenClaw that keeps prompt packs smaller, trust tiers visible, and receipts explicit for why a memory was included, excluded, or left fail-open.

Public messaging baseline¶

Problem¶

OpenClaw memory should stay trustworthy under pressure, not quietly import stale or hostile context.

Product approach¶

Trust-aware context packing: smaller cited packs, visible trust tiers, and inspectable selection receipts.

Proof surface¶

Run the synthetic before/after proof and show that a trust policy can exclude quarantined rows while keeping receipts intact.

Adoption guidance¶

Adopt sidecar-first; promote to mem-engine only when hybrid recall/policy controls are needed.

Claims guardrails¶

Keep these claims true: - pack exists today - --trace exists today - --pack-trust-policy exclude_quarantined_fail_open exists today - policy_surface and lifecycle_shadow receipts exist today - the proof artifact is synthetic and reproducible - graph provenance policy exists as a compatible extra surface, but it is not required for the basic proof

Avoid these claims: - “solves memory forever” - “fully blocks all bad memory automatically” - “ships full provenance URLs for every citation today” - “replaces all native OpenClaw memory by default” - “graph/reference is the universal schema for everything”

Boundary rules for outward copy: - KOL/GTM is linked but separately governed; never merge control-lane authority into product copy. - Query-plane default and action-plane write-gated posture must remain explicit. - For publicly released materials, internal narrative and external narrative must remain clearly separated; do not publish internal framing formulas or backstage positioning language as outward-facing copy.

Release-ready snippet¶

openclaw-mem is positioned around trust-aware context packing, not generic memory storage.

Long-running agents do not only forget. They also admit stale, untrusted, or hostile context and quietly drag it into future prompts. openclaw-mem gives OpenClaw a local-first way to build smaller, cited prompt packs with explicit trust tiers and trace receipts for inclusion/exclusion/fail-open decisions.

The canonical proof uses the same query against the same DB before/after trust policy: a quarantined row drops out, a trusted row takes its place, the pack gets smaller, and the receipts stay intact.

Proof: - docs/showcase/trust-aware-context-pack-proof.md - docs/showcase/artifacts/trust-aware-context-pack.metrics.json

Short release blurb¶

Trust-aware context packing for OpenClaw: smaller prompt packs, explicit trust tiers, and receipts for why memory entered the pack.

GitHub-facing surfaces¶

Repo description / pin-safe line¶

Trust-aware context packing for OpenClaw — pack only what the agent should trust, with provenance, receipts, and local-first recall.

Ultra-short line¶

Trust-aware context packing for OpenClaw.

Trust-aware context packing for OpenClaw

Pack only what the agent should trust

X / thread-ready copy¶

Single-post version¶

openclaw-mem's core value proposition is now much cleaner:

trust-aware context packing for OpenClaw - smaller prompt packs - explicit trust tiers - recordRef citations + trace receipts - lower chance that stale / hostile context quietly re-enters future prompts

The proof artifact is the key: same query, same DB, different trust policy — quarantined row drops out, trusted row takes its place, bundle gets smaller.

Thread version¶

Most “AI memory” products pitch storage. The nastier production problem is admission: stale, untrusted, or hostile content quietly enters future prompts.
openclaw-mem narrows around that: trust-aware context packing for OpenClaw.
Baseline is local-first and inspectable: SQLite + JSON receipts + search → timeline → get → pack.
The packing surface: --trace --pack-trust-policy exclude_quarantined_fail_open policy_surface lifecycle_shadow
Canonical proof: same query, same DB, same limit. Turn on trust policy and the quarantined row is excluded, a trusted row replaces it, and the pack shrinks.
Whole value proposition: smaller / safer prompt packs with receipts — not “store everything and pray.”

30-second opener¶

openclaw-mem is about trust-aware context packing, not generic memory storage. The problem is not only forgetting. It is letting stale or hostile text become durable memory and then quietly re-enter prompts later.

3-beat demo flow¶

Show the ungated pack
run the proof fixture without trust policy
point out that a quarantined row still enters the pack because it text-matches
Turn on trust policy
rerun with --pack-trust-policy exclude_quarantined_fail_open
show that the quarantined row is excluded with an explicit reason
Close on receipts
show trace, trust_policy, policy_surface, and lifecycle_shadow
emphasize selection changed with logs; memory was not silently mutated

Closing line¶

The product promise is not “more memory.” It is smaller, safer prompt packs with trust tiers and receipts.

Demo links to keep handy¶

Proof doc: docs/showcase/trust-aware-context-pack-proof.md
Metrics JSON: docs/showcase/artifacts/trust-aware-context-pack.metrics.json
Raw fixture: docs/showcase/artifacts/trust-aware-context-pack.synthetic.jsonl
Companion demo: docs/showcase/inside-out-demo.md
Operator lock checklist: docs/launch/proof-first-relaunch-checklist.md
Release-surface proof pack: docs/launch/release-surface-proof-pack-v0.md