Please check the build logs for more information.
See Builds for ideas on how to fix a failed build, or Metadata for how to configure docs.rs builds.
If you believe this is docs.rs' fault, open an issue.
rullama-memory
Tiered hot/warm/cold agent memory orchestration for the rullama agent framework.
What this crate is
The schema layer — the five tier stores (MessageStore,
SummaryStore, FactStore, MentalModelStore, TierMetadataStore)
plus the shared tier_types (MemoryTier, MemoryAuthority,
TierMetadata, MessageSummary, KeyFact, FactType) — lives in
rullama-stores.
This crate adds:
TieredMemory— multi-factor adaptive search across all four tiers (similarity × recency × importance), with promotion / demotion of entries when access patterns change.dream(feature-gated) — offline consolidation engine that summarises hot-tier messages into warm-tier summaries, extracts cold-tier facts, and demotes by retention score.
Why a separate crate
rullama-stores is schema only — table definitions and CRUD. The
orchestration (TieredMemory) and consolidation (dream) live here
because they are engines, not schema: search policy, scoring
weights, write authority gating (CanonicalWriteToken), demotion
heuristics, summarisation pipelines.
A consumer that only needs to write rows directly to one tier (e.g. a
test) depends on rullama-stores alone. A consumer that wants
multi-tier adaptive search or offline consolidation pulls
rullama-memory, which re-exports the schema types it operates
over.
Feature flags
dream— offline consolidation engine. Pullsfuturesfor the async cycle plumbing.telemetry(impliesdream) — wire dream'swith_analyticshook torullama-telemetry'sAnalyticsCollector.
License
MIT OR Apache-2.0