Compounding Context OS roadmap

Status: PUBLIC ROADMAP / DESIGN NOTE

openclaw-mem is evolving from a memory sidecar into the memory and context layer for a compounding agent operating system. The goal is not to make a larger chat transcript. The goal is to maintain a small, governed, explainable context supply chain that improves as agents work.

Product stance

Keep the stable split:

• Store captures durable records with provenance, trust, importance, and rollback-friendly receipts.
• Pack compiles a bounded ContextPack for the current task. It owns the injection contract.
• Observe keeps artifacts, traces, and review receipts inspectable without flooding the live prompt.

Graphs, docs memory, GBrain-style sidecars, and lifecycle jobs may enrich this loop, but they should not become competing sources of truth.

Why this matters

Long-running agents fail when memory becomes a pile of stale snippets. A useful context system needs to answer four questions on every turn:

1. What should be remembered?
2. What should be packed right now?
3. What should be archived, quarantined, or rechecked?
4. What receipt proves why that choice was made?

The compounding path is therefore a governance loop, not only a retrieval loop.

Capability map

Capability	Home surface	First-class contract
Durable record capture	Store	provenance, trust, importance, citations
Task context assembly	Pack	openclaw-mem.context-pack.v1
Memory lifecycle review	Store + Observe	review-only steward receipts
Source ingestion	Store	public-safe candidate records and entity hints
Skill/workflow learning	Observe	lifecycle receipts before apply
Active execution lines	External controller	goal/status/stop-loss receipts that Pack can cite

Memory steward review

The first additive slice is a deterministic, review-only steward contract:

• classify candidate records as promotion, keep-observed, archive/ignore, or quarantine candidates;
• surface public-safety review markers before public docs are written;
• emit no side effects;
• require a separate checkpointed apply path for any mutation.

This keeps the system safe-by-default: the steward can recommend changes, while Store/Pack/Observe remain auditable and rollbackable.

Design rules

1. Review before apply — steward output is advisory until a separate apply path checkpoints, mutates, and verifies.
2. Pack remains bounded — lifecycle jobs may influence ranking, but they do not justify dumping more raw memory into the prompt.
3. Trust is separate from usefulness — frequent retrieval must not promote untrusted content to trusted content.
4. Public docs stay public-safe — examples should avoid private paths, identities, channel IDs, tokens, and local operator details.
5. Receipts beat vibes — every automation loop should leave machine-readable evidence plus a human-readable summary when operators need debugging context.

Shipped public-safe slices

• openclaw-mem steward review --file <json-or-jsonl> reviews candidate records without mutating storage.
• openclaw-mem ingest-review source --file <text-or-markdown> turns a source into candidate records, entity hints, follow-up actions, and risk markers.
• openclaw-mem active-line pack --file <receipt.json> converts an active-line or goal receipt into a small ContextPack-compatible fragment.

Suggested next slices

1. Connect steward receipts to pack --trace so included records can refresh lifecycle metadata through a governed apply path.
2. Add richer synthetic ingestion fixtures for articles, transcripts, PDFs, and repository notes.
3. Let Pack accept active-line context fragments directly as protected L0 inputs without loading a whole operations log.

Non-goals

• Replacing the operator's canonical memory backend by force.
• Auto-mutating trusted memory from untrusted input.
• Making graph or sidecar experiments the source of durable truth.
• Treating a human-readable report as a machine gate unless that gate is separately specified and tested.