Compiled from
raw/codebase-snapshots/supabase-schema-2026-04-09.md(mm_memory_items,mm_doc_records)
Lexery - Memory and Documents
Why This Page Matters
Пам’ять у Lexery — це не дрібна embellishment-фіча. У legacy architecture вона була однією з центральних систем, а в current repo її сліди вже існують у runtime, contracts, outbox patterns і document surfaces.
Legacy Vision
Memory Manager
- Full logical scheme in bridge repo
plan.mdincludes: users, profiles, cases, conversations, messages, summaries, memory items, run snapshots. - Scopes: conversation, case/project, user-global.
- Types: preferences, facts, constraints, glossary, notes.
- Flow: proposed → confirmed/rejected.
ContextPack
- Memory was designed to enter the runtime in bounded blocks: profile, case, rolling summary, memory items, recent messages, retrieved history.
Case digest / rolling summaries
- Memory was expected to produce structured case-level and conversation-level summaries for lawyer workflow continuity.
Current Observed Surfaces
In apps/brain
mm/mm/doc/- memory tools
- outbox-related stabilization issues in Linear
In contracts and product surfaces
- Attachments already appear in shared request schemas.
- Portal already has attachments UI.
- Gateway and R2 attachment overflow already exist in Brain docs/readmes.
Document Story
Observed
- Portal includes attachments panel and file preview/UI.
- API includes presigned upload endpoint.
- Brain README describes R2 overflow attachments under
runs/{tenant_id}/{run_id}/attachments/. - MM docs are explicitly mentioned as a runtime surface.
Inferred
- Documents are intended to become a first-class retrieval/memory layer, not just chat add-ons.
Current Maturity Reading
Observed: the architectural vision for memory is far ahead of what is clearly exposed as stable product behavior today.Observed: memory/outbox correctness is still an active stabilization topic in Linear.Inferred: Lexery sees durable context as one of the key differentiators for professional legal use.
Tension
Strong vision
- structured memory
- semantic indexing
- scoped recall
- doc-aware context
- post-run extraction
Strong implementation risk
- cross-tenant correctness
- outbox integrity
- ordering and idempotency
- privacy and retention
- keeping memory useful rather than noisy
Best Synthesis
Memory in Lexery is best understood as the second brain behind the Brain:
- retrieval answers “what law is relevant?”
- memory answers “what from this user/case/project history still matters?”
The vision is already clear. The implementation is partially present and still being hardened.
See Also
- Lexery - Pipeline Health Dashboard
- Lexery - Brain Architecture
- Lexery - Product Surface
- Lexery - Linear Roadmap
- Lexery - Open Questions and Drift
- Lexery - Storage Topology
- Lexery - Provider Topology
- Lexery - U9 Assemble
- Lexery - Run Lifecycle
- Lexery - Unknowns Queue
- Lexery - Retrieval, LLDBI, DocList
- Lexery - LLDBI Surface
- Lexery - Import Proposal Loop