Three defects the first sweep left behind, each caught by a rule the guard did not yet have. Statuses. RULES.md gives three; the tree had 25 values, and the free-text ones were carrying progress. "Implemented (wiring landed)" and "Ratified (rev. 2e). Rev. 2e adds the OpenTaskByPR lookup. Rev. 2d revised section 10..." were both changelogs living in the one field the other rules exempt. Every PRD now declares Draft, Approved or Finalised and nothing else; five that declared nothing now declare one. A revision number moves to **Revision:**, a pointer to the authoritative document to **Authority:**, and what a past revision changed to git. CLAUDE.md's citation of state-store-contract rev. 2e still resolves. Stale cross-references. "Supersedes the 2026-09-03 open question", "this bullet previously said", "an earlier draft of this document claimed" all point at text the reader cannot see, because it was replaced. A supersede marker survives only when it names a file, a section or a symbol that can be found. Temporal hedges. 100 uses of today, currently and right now across 31 files. Each dates its sentence to the day it was written: "[notify] today is one webhook URL, read in one place" stops being true the moment this design lands, while still reading as the premise it rests on. The header-field exemption was also too wide. "**Date:** 2026-09-03, Channel resolved 2026-09-22" smuggled a dated status into the one line the rules did not read; the exemption is now a Date holding only a date and a Status holding only one of the three.
9.6 KiB
Persistent session-memory curator (reference implementation) -- decision
Status: Approved
Date: 2026-09-02
Beads issue: archie-core-1786637499114
Parent epic: archie-core-1786637500725 (curator engine)
Depends on: archie-core-1786637499636 (bind curators to memory
engines, shipped) and the domain/memory engine family (shipped)
Decision
Each pass: find sessions with activity since the last pass, ask the
model to extract a short list of durable observations from each one's
recent messages, and write each through the curator's bound memory
engine. Extraction is scoped to one session -- Observation.Identity
is the session ID, not a user or cross-session identity. No
consolidation across sessions, no forgetting, no rewriting a prior
extraction: every pass only adds.
This is the second contract consumer the epic asks for, and the one that
exercises the curator/memory-engine binding
(archie-core-1786637499636) for real, unlike the skill curator, which
declares no memory engine at all.
Why session-scoped, not user- or agent-wide
docs/architecture/migration-decisions.md#5 lists the four required
memory scopes (global, agent-wide, user-wide, agent-user relationship)
as still open -- there is no decided mapping yet from a chat session
to a durable identity that would outlive that session (a Telegram chat
ID is not necessarily "the same person" as a web session; per-platform
identity resolution does not exist as a memory concern). Choosing
session-as-identity sidesteps a decision this issue has no business
making unilaterally: it is the narrowest scope that is unambiguous
(gateway.SessionContext.SessionID already exists, is already unique,
and already has no cross-platform identity claim attached to it), and
it is forward-compatible -- widening to a resolved user identity later
is a change to what Identity is populated with, not to the
extraction or write path.
Why a new curator.ConversationSource, not gateway.SessionStore directly
internal/domain/curator is a domain package; internal/gateway is not
one of internal/domain/*'s dependencies by convention, and
gateway.SessionStore's full CRUD surface (branch, fork, replace,
search) is far more than a curator needs or should be able to reach --
the same reasoning ToolBuilder/SkillStore/MemoryEngineSource
already established for every other host service this package depends
on. curator.ConversationSource is a two-method read-only interface;
internal/app/archied adapts the real gateway.SessionStore to it, the
same shape MemoryEngineSource already uses for
domain/memory.Registry.
// ConversationSource is read-only conversation history, narrowed to
// what a curator needs to review recent activity -- never the full
// session CRUD/branch/search surface a chat gateway exposes.
type ConversationSource interface {
// RecentSessions returns every session touched at or after since,
// newest first.
RecentSessions(ctx context.Context, since time.Time) ([]SessionSummary, error)
// Messages returns the most recent n messages of one session, in
// chronological order.
Messages(ctx context.Context, sessionID string, n int) ([]ConversationMessage, error)
}
type SessionSummary struct {
ID string
LastActive time.Time
}
type ConversationMessage struct {
Role string // "user" or "assistant"
Content string
At time.Time
}
Manifest gains Conversations bool, gated exactly like Skills:
Registry.validateDeclared requires a ConversationSource on the host
when declared, Registry.filter clears it from a curator's view when
not. Role is derived the same way gateway.compressTurnHistory already
does (message.From == botUser -> "assistant", else "user") --
copied logic, not shared code, since the source type
(gateway.Message) is exactly the thing this package must not import.
Extraction shape
Pass:
Conversations.RecentSessions(ctx, effectiveSince(in.Since))-- see below for the zero-Sincedefault.- For each session,
Conversations.Messages(ctx, id, 40)-- the last 40 messages, not the full history. A session accumulates far more history than a skill file accumulates content; unlike the skill curator's "re-read everything, it's cheap," a full-history read here is not cheap and is not needed -- durable facts worth keeping surface in recent exchanges, and a 40-message tail is deliberately generous against the model's own context rather than tuned tightly, since getting this number exactly right is not this issue's job. - One model call per session: the message tail, plus an instruction to
return a JSON array of short strings, each one fact/preference/
correction worth keeping, or
[]if nothing is. Capped at 5 items per session per pass -- a session that produced more than 5 durable facts in one pass is the extraction prompt's problem to get more precise about, not a limit this curator should quietly lift. - A response that isn't valid JSON, or that fails to parse, is treated as "nothing extracted" for that session, not a pass failure -- the next pass tries again, and this is a curator, not a build gate.
- Each extracted string becomes one
WritethroughRegistrar.MemoryEngines.Get("builtin")(Observation{Identity: sessionID, Kind: "note", Content: text}). OneActionper session summarizes the count and lists the session ID; a session with nothing extracted gets noAction, matching the skill curator's "a pass with nothing to report is not an error, and not every subject needs an entry every time."
effectiveSince: PassInput.Since zero (first pass, or the runtime
declining to name a horizon) resolves to 24 hours before now, not "the
beginning of time." Extracting from a session's entire history on the
very first pass is exactly the kind of unbounded-cost-on-cold-start
mistake Check exists to keep cheap, and 24h is long enough to catch
what a daily check-in cadence would have caught anyway.
Check: len(RecentSessions(ctx, effectiveSince(zero-Since))) > 0 --
cheap (no model call), and answers "is there anything a pass would find"
honestly rather than always returning true.
What we deliberately do NOT do
- No cross-session or cross-user consolidation.
Identityis the session ID; two sessions from the same human get two disjoint memory scopes until the open identity-resolution decision inmigration-decisions.md#5is made. - No re-extraction guard, no deduplication. A fact mentioned in two
overlapping passes may be written twice.
domain/memory.Queryranking/relevance is a read-side concern for whatever later consumes these observations, not this curator's job to solve by inventing a dedup index. - No
Forget. Nothing here decides a memory is stale; that is a different, harder judgment call than "is this worth keeping," and nothing reads these observations back into a chat turn to notice if one turned out wrong. - No tool declarations.
Manifest.Toolsis empty; the model call is a plain completion (ChatRequest.Toolsleft nil), not agentic tool-calling -- the extraction task does not need tools, and declaring none is what keepsRegistrar.Toolscorrectly absent from this curator's view. - No feeding extracted observations back into the live system
prompt.
internal/memory.Manager's existing frozen-snapshot mechanism is untouched, same disclaimer as the memory-engine-contract and builtin-engine issues before this one: the newdomain/memorypath and the legacy live path run in parallel by design. Reading these observations back into a chat turn is later, separate work.
Acceptance criteria
Restated from the issue against this shape:
- Registers and runs through
curator.Registrywith no special-casing -- the existingRegister/Runtimepath, no new registry method. - Writes through the memory engine contract, never a concrete store:
every write goes through
Registrar.MemoryEngines.Get(name).Write, resolved by name at call time exactly like the skill curator resolves nothing store-specific either. - Extraction runs off the hot path: the model call happens inside
Pass, which the curator runtime already runs off a chat turn's own goroutine (existing runtime guarantee, not something this issue builds). - What it wrote and why is inspectable: one
Actionper session with a count and the session ID, through the existing curator activity tracker -- no new observability surface. - Honours the trigger-accounting invariant:
Manifest.OnInput = truewakes it onevents.KindTurnCompletedvia the existingWakeOnPrimaryInputwiring, which already excludes every curator-derived event kind from the forwarded set -- this curator's own writes cannot reschedule it, with no new code needed for that guarantee.
Files this change adds
internal/domain/curator/registrar.go--ConversationSource,SessionSummary,ConversationMessage, newConversationsfieldinternal/domain/curator/registry.go--Conversationsgated likeSkillsinvalidateDeclared/filterinternal/domain/curator/registry_test.go-- table cases + fakes for the new field, mirroring the existingSkills/MemoryEnginescasesinternal/infrastructure/sessioncurator/adapter.go-- wrapsgateway.SessionStoreascurator.ConversationSourceinternal/infrastructure/sessioncurator/adapter_test.gointernal/infrastructure/sessioncurator/sessioncurator.go-- theCuratorEngineinternal/infrastructure/sessioncurator/sessioncurator_test.gointernal/app/archied/bootstrap.go-- register it insetupCurators, wired tob.chatSessionStoreandb.cfg.BotUser