Plan 0039
One memory, queryable across the whole world
Agents write observations and claims into one Datalog memory (weave); consumers query the world instead of grepping hosts. Fleet rollups like "top things agents are irritated by, above a threshold" are saved projections, not features.
- Status
- Input wanted RFC
- Ambition
- 8 one query language over memory, telemetry, and live system state
- Impact
- 7 cross-cutting questions about the fleet become queries instead of archaeology
- Effort
- 5 weave substrate is executed; the delta is adapters, projections, and a read surface
- Risk
- 4 rollups over noisy claims can mislead; projections are derived and cheap to fix
- Maturity
- 2 substrate landed in weave; no consumer rollup exists yet
- Leverage
- 8 every future dashboard or fleet question is a saved projection, not a subsystem
- Taste
- 7 deciding what becomes a claim and which rollups earn a surface is the judgment
- Authors
andrewgazelka- Created
- Updated
- Tracking issue
- -
- Supersedes
- -
- Superseded by
- -
Summary
Agents write observations and claims into one memory: weave, Datalog facts with provenance envelopes (weave#281, executed). Consumers then query the world instead of grepping hosts. “The world” is wider than what agents journal: pull and push adapters make live state facts by evaluation time (weave docs/weave-datalog-over-everything.html), so one query language spans memory, telemetry, and system state.
Motivation
Cross-cutting questions (“what are agents most irritated by this week, across the cluster?”) have no query today; the evidence is scattered over transcripts, issues, and per-host logs. Once every irritation, blocker, and gotcha is a stated claim with provenance, that question is a saved projection: group by subject, count distinct stating sessions, filter to current, apply a threshold, rank.
Detailed design
- Substrate: weave agent memory plus the datalog-over-everything adapter model for non-journal sources. Nothing here reopens those designs.
- Consumer rollups are projections a consumer registers, not server features. First dashboards: top irritations above a threshold, recurring blockers, most-hit gotchas, per-repo friction.
- ix-mcp stays the write choke point; the delta this plan owns is the read surface where projections render for humans (site page or TUI panel).
- Plan 0021 (filesystem as database) becomes one adapter in this frame: files are entities, xattrs components, its index a pull-through source.
- Plans 0025 (agent gossip), 0026 (shared brain), 0028 (cross-agent memory), and 0030 (knowledge graph) name write paths and stores; this plan names the query substrate they can share: facts in Datalog, rollups as saved projections. 0025’s irritation clusters are the first projection this plan wants rendered.
Open questions
- Which rollups earn a rendered dashboard versus staying ad hoc queries?
- Threshold semantics: absolute counts, distinct sessions, or recency-decayed?
Drafted by Claude Code (Fable); comments wanted, hence the RFC flag.