Introduce curators as their own engine family: long-running agent loops that reason over memory and perform small maintenance tasks, rather than one hardcoded background job. A curator is defined by a check-in interval, a tool set, and (once the memory engine family lands) an attached memory engine. It runs persistently while there is input and backs off to a cooldown when idle. Ship two reference curators -- a skill curator and a persistent-session-memory curator -- so the interface is proven against two consumers rather than designed around one.
Design
The design splits a cheap hot-path extractor from an expensive sampled reasoner, so agentic cost is never paid per message. Sampling chooses which memories deserve attention via surprisal-based scoring, implemented over embeddings through interchangeable strategies including a k-NN graph random walk. Archie has no embedding capability today, so any surprisal-style sampling is gated behind adding one.
Acceptance Criteria
Curator is a conforming engine family per ARCHITECTURE.md 'Plugin engine rule (strict)': typed contract, owning registry/manager, start/health/stop, narrow typed host access. Curators are registered and configured declaratively, not hardcoded. The runtime supports persistent operation under load and cooldown when idle, with intervals configurable per curator. Both reference curators run against the shipped defaults. Curator activity is observable: what ran, when, what it changed, and why. A curator failure is isolated and never takes down the daemon or a chat turn.
Notes
Status as of 2026-09-02: all 7 real children closed (contract/registry,
runtime, observability, trigger-accounting, memory-engine binding, and
both reference curators -- skill and session-memory).
One epic AC clause remains genuinely unmet: "curators are registered and
configured declaratively, not hardcoded." Both reference curators are
registered via direct Register() calls in
internal/app/archied/bootstrap.go:setupCurators, with hardcoded
intervals (skillcurator.DefaultInterval, sessioncurator.DefaultInterval)
-- no config surface exists to enable/disable a curator or tune its
interval/cooldown without a code change and rebuild. Leaving this epic
open on that basis rather than closing it on a partial match; a
follow-up issue for declarative curator config (mirroring how
archie-core-1786637490043 did this for memory.engine) would close it.
Close reason
All children completed
Imported from beads archie-core-1786637500725-414-9d66a52b · GitHub #435 · type epic · priority P2 · created 2026-08-08 · closed 2026-09-02
Introduce curators as their own engine family: long-running agent loops that reason over memory and perform small maintenance tasks, rather than one hardcoded background job. A curator is defined by a check-in interval, a tool set, and (once the memory engine family lands) an attached memory engine. It runs persistently while there is input and backs off to a cooldown when idle. Ship two reference curators -- a skill curator and a persistent-session-memory curator -- so the interface is proven against two consumers rather than designed around one.
### Design
The design splits a cheap hot-path extractor from an expensive sampled reasoner, so agentic cost is never paid per message. Sampling chooses which memories deserve attention via surprisal-based scoring, implemented over embeddings through interchangeable strategies including a k-NN graph random walk. Archie has no embedding capability today, so any surprisal-style sampling is gated behind adding one.
### Acceptance Criteria
Curator is a conforming engine family per ARCHITECTURE.md 'Plugin engine rule (strict)': typed contract, owning registry/manager, start/health/stop, narrow typed host access. Curators are registered and configured declaratively, not hardcoded. The runtime supports persistent operation under load and cooldown when idle, with intervals configurable per curator. Both reference curators run against the shipped defaults. Curator activity is observable: what ran, when, what it changed, and why. A curator failure is isolated and never takes down the daemon or a chat turn.
### Notes
Status as of 2026-09-02: all 7 real children closed (contract/registry,
runtime, observability, trigger-accounting, memory-engine binding, and
both reference curators -- skill and session-memory).
One epic AC clause remains genuinely unmet: "curators are registered and
configured declaratively, not hardcoded." Both reference curators are
registered via direct Register() calls in
internal/app/archied/bootstrap.go:setupCurators, with hardcoded
intervals (skillcurator.DefaultInterval, sessioncurator.DefaultInterval)
-- no config surface exists to enable/disable a curator or tune its
interval/cooldown without a code change and rebuild. Leaving this epic
open on that basis rather than closing it on a partial match; a
follow-up issue for declarative curator config (mirroring how
archie-core-1786637490043 did this for memory.engine) would close it.
### Close reason
All children completed
---
_Imported from beads `archie-core-1786637500725-414-9d66a52b` · GitHub #435 · type epic · priority P2 · created 2026-08-08 · closed 2026-09-02_
sam
added the mediumepic labels 2026-10-27 13:38:23 +00:00
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Introduce curators as their own engine family: long-running agent loops that reason over memory and perform small maintenance tasks, rather than one hardcoded background job. A curator is defined by a check-in interval, a tool set, and (once the memory engine family lands) an attached memory engine. It runs persistently while there is input and backs off to a cooldown when idle. Ship two reference curators -- a skill curator and a persistent-session-memory curator -- so the interface is proven against two consumers rather than designed around one.
Design
The design splits a cheap hot-path extractor from an expensive sampled reasoner, so agentic cost is never paid per message. Sampling chooses which memories deserve attention via surprisal-based scoring, implemented over embeddings through interchangeable strategies including a k-NN graph random walk. Archie has no embedding capability today, so any surprisal-style sampling is gated behind adding one.
Acceptance Criteria
Curator is a conforming engine family per ARCHITECTURE.md 'Plugin engine rule (strict)': typed contract, owning registry/manager, start/health/stop, narrow typed host access. Curators are registered and configured declaratively, not hardcoded. The runtime supports persistent operation under load and cooldown when idle, with intervals configurable per curator. Both reference curators run against the shipped defaults. Curator activity is observable: what ran, when, what it changed, and why. A curator failure is isolated and never takes down the daemon or a chat turn.
Notes
Status as of 2026-09-02: all 7 real children closed (contract/registry,
runtime, observability, trigger-accounting, memory-engine binding, and
both reference curators -- skill and session-memory).
One epic AC clause remains genuinely unmet: "curators are registered and
configured declaratively, not hardcoded." Both reference curators are
registered via direct Register() calls in
internal/app/archied/bootstrap.go:setupCurators, with hardcoded
intervals (skillcurator.DefaultInterval, sessioncurator.DefaultInterval)
-- no config surface exists to enable/disable a curator or tune its
interval/cooldown without a code change and rebuild. Leaving this epic
open on that basis rather than closing it on a partial match; a
follow-up issue for declarative curator config (mirroring how
archie-core-1786637490043 did this for memory.engine) would close it.
Close reason
All children completed
Imported from beads
archie-core-1786637500725-414-9d66a52b· GitHub #435 · type epic · priority P2 · created 2026-08-08 · closed 2026-09-02