Experience retrieval and memory strategies
1. Three sources with different jobs
P Experience retrieval is an LLM-guided skill, not a binary named “experience search.” Its software queries return candidates; the agent chooses sources, interprets similarity and checks current applicability. A future executable helper can automate a proven substep, but cannot replace this strategy with a keyword match.
| Source | What it answers | Freshness and authority |
|---|---|---|
| Current case registry and arrival identities | Is this the same source request/revision? Is related work already open, waiting or completed? | Live case state. Deterministic source identity handles delivery retries; semantic similarity alone does not merge requests. |
| Reviewed experience summaries | What outcome, route, mechanism or effort was established in a comparable case? | Versioned summary with source case/origin, owner, observed result, limits and correction/supersession reference. A proposed fix is not an executed fix. |
| Domain memory and methods | How does this system or kind of work behave, and how should it be approached? | Owned articles and skill revisions, with source/version scope. Current system and consuming code outrank stale memory. |
P These are complementary reads. An open case does not need to wait for precipitation before it is discoverable in the case registry. A closed ticket is not automatically a verified method. A method is not a reusable permission.
2. Strategy is selected by the requesting agent
| Strategy | Starts with | Looks for | What the agent must decide |
|---|---|---|---|
| Root / classification | Authenticated customer and request identity, current case registry, cross-domain routing summaries | Duplicate source delivery, related open work, past request types, accepted/rejected routes, owner and contractual disposition | Same request or independent occurrence? Data work, Configuration, Development, Assessment & Offers, or an existing wait? A similar title does not establish a mechanism. |
| Data investigation / resolution | The symptom domain's methods and confirmed experience, current target and lifecycle | Mechanism and failure discriminator, required checks, actual repair outcome and residual effects | Does the known pattern fit current evidence, or must investigation start from the symptom? SD-1598's absent register row is different from a stale PDF or an uncreated renewal. |
| Configuration | Normalization/stage memory, current product baseline, source set and prior design decisions | Document traps, matched product/template variants, factor/cover/selector relationships, prior semantic assumptions | Which source means what, which stages are needed, and which old facts must be re-read? Never copy old tariff cells or call compiled templates correctly selected. |
| Development | Current source/branch set and repository memory, comparable changes and regressions | Mechanism, appropriate seam, consuming interfaces, test and release obligations | Does the old change apply to this actual lineage and consumer set? A remembered path or commit is not current serving proof. |
| Effort assessment | Agreed scope, completed comparable work, previous estimates and their actual results | Work breakdown, actual active effort, waiting/dependencies, changed scope and estimation error | What is comparable, what is new, and what range/uncertainty is justified? Historical effort does not supply today's price or committed date. |
P Each profile pins a default strategy. The task can broaden its search when evidence requires it, recording why. Domain indexes remain discoverable; an agent does not load every article into its prompt. A consulted agent starts from its own domain and scoped brief. Consult replies are evidence, not grants or adopted conclusions.
3. What Support root actually retrieves
P Root first resolves the arrival's canonical origin and verified tenant. It queries authorised current cases for the same origin/object/request relationship and reads routing summaries across the module domains. Summaries carry request type, symptom tags, product/system, owner, chosen route, actual disposition, source case, result reference, current revision and validity limits. They omit raw personal data and the full technical transcript.
P An example is “printing request → Data printing investigation → external template owner → verified document result,” not “apply the SQL from ticket X.” Another is “new shared numeric parsing behaviour → effort assessment and offer,” not “start a Source implementer.” Root can consult the owning domain if classification depends on a technical distinction. The selected Data or Configuration agent then retrieves the deeper method and performs current checks.
P Case visibility remains tenant/scope-bound even though sanitised domain knowledge is broadly readable. Cross-customer experience may expose only the approved general pattern; current case bodies and identity handles remain scoped. A new independent request obtains new plan/permissions and records its own applicability evidence.
4. Indexing without copying every memory everywhere
P Keep one canonical owned article/skill and a versioned search/routing projection referencing it. Domain owners do not replicate their entire memories into classification memory. The root index contains discovery metadata and routing synopses; detailed content is fetched from the owning domain when needed.
P Search indexes are derived lookup structures, not authored memory nodes. The projector does not rewrite another domain's AGENTS.md or invent a new narrative. Any textual synopsis is part of the owner-reviewed experience; the projector copies its approved fields and canonical references into the searchable projection.
- A verified case/stage result creates a sanitised experience proposal linked to the canonical source request and actual execution/proof records.
- The experience curator and relevant domain owner review the claimed route/mechanism/method. Rejected or unverified outcomes remain labelled; they cannot be promoted as successful execution.
- Publication commits the active revision and an outbox event together. A deterministic index projector updates the owning-domain index and routing synopsis, referencing the same canonical revision.
- The projection records its source revision/watermark. Corrections and retirement invalidate/supersede indexed entries; they do not leave the old assertion silently searchable as current truth.
- A retrieval request records the selected strategy, index watermark, candidates, actual reads and applicability verdict in case papers / memory-use records. A lagging index prompts a scoped canonical lookup or refresh; zero indexed hits is not proof of absence.
P Use the existing case, arrival, experience, memory-domain/article, papers, memory-use and outbox records (Data Model, Agent Runtime). Search-specific metadata is a versioned projection/strategy contract; no duplicate “all memories” store or new autonomous memory-writing agent is required. Case registry state updates independently of post-result learning.
5. Retrieval contract for implementation
P Input: strategy id/version; requesting profile/domain and purpose; tenant/case access scope; current objective/symptom and canonical source identity; current product/system/environment/source versions; required trust/freshness and result budget. The backend derives access from identity and case grants, not model-supplied scope expansion.
P Output: candidate kind and owner, canonical ref/revision, source case/origin, observed outcome and proof level, similarity reason, applicability limits, index watermark and known missing/lagging sources. The LLM returns a cited choice/rejection and the fresh discriminating checks it needs. The runtime stores that decision; a search score does not choose or execute a remedy.
P Acceptance cases: duplicate delivery attaches once; similar independent cases remain separate; a rejected route informs classification without becoming a successful fix; stale schema/version causes revalidation; a retired article disappears from current candidates; index lag is visible; a family with uncreated renewals does not reuse the missing-row repair; estimate comparisons separate active effort from owner waiting. The existing retrieval baseline measures retrieval quality; it is not the production strategy implementation.