Repository navigation
Replies: 2 comments
|
Thanks for breaking this out of #1254 into a focused RFC — I checked the claims against v11.12.0 and current What's missing (confirmed)
What's shipped from #1047 (also confirmed)
Prior art (confirmed, one link fix needed)
Direction from my side Before opening the first PR, one thing I'd like in each: a short doc of what the window tools return (payload shape + provenance), since |
|
One short scope-clarification in the same spirit: this trilogy is about background enrichment, not inter-agent messaging — and messaging is already an established, zero-build pattern here. The README's "Multi-Agent Cluster with Shared Memory" section (@jeremykoerber, issue #591) shows agents exchanging messages purely via sentinel tags ( |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Moving #1254 here for discussion, as we're routing design RFCs to Discussions.
Context
This is the one piece of the old #1047 (Self-Service Memory Intelligence) that never landed on either side. Everything else from #1047 shipped: observation store, belief pipeline (
BeliefService+ migration 012), the dream-inspired consolidator,get_bootstrap_profile,commit_session_legacy. What's still missing is the background enrichment trilogy that ran on the consolidation scheduler in the fork.This is a focused proposal — not the whole #1047 (that RFC was too broad), just the three background jobs.
What's missing (verified against v11.12.0)
memory_facts/memory_gapstools intools/registry.py013/014/015fact_extraction/gap_detection/feedback_recalculationjobs inconsolidation/scheduler.pyProposal (three independent background jobs, all opt-in)
memory_factswindow tool for inspection). Migration 013.memory_gapswindow). Migration 014.Each is independent, opt-in via env, and runs in the background via the existing scheduler — the service works without them. They enrich the memory the agent reads.
Why now / value
These are the highest-value pieces still parked. The fork has a working prototype (
2e1978d5+ tests). Before writing any code I'd like your read on scope and whether this fits the post-v11 batch-review ground rule — happy to split into three separate PRs (one per job + migration) so each stays reviewable on its own.Prior art: fork RFC
docs/rfc/rfc-self-service-memory-intelligence.md§3; the Codeberg issue #121 (belief derivation) carried the belief half.All reactions