ContextLattice turns scattered chats, tool outputs, decisions, skills, checkpoints, and evidence into local-first memory infrastructure: fast enough for the next answer, durable enough for next week, and deep enough to keep learning as the system expands.
Workspaces help humans steer models. Agent harnesses let models act. ContextLattice gives both shared infrastructure for remembering: one local-first write path, one retrieval contract, one CLI-native workflow, one Skills Index, prompt-ready session packages, and traceable run cards that show which evidence shaped the work.
topic_rollups keep the recent project shape close to the agent so startup recall is fast and cheap.
qdrant gives the base local stack a vector-native lane without changing the agent contract.
Every checkpoint, decision, and evidence note enters through a stable write path before it fans out anywhere else.
Full/operator stacks add pgvector, raw ledger, graph edges, Letta, memory-bank acceleration, observability, and slower async recall without asking agents to learn a new interface.
Retrieve the sharpest slice of context instead of dragging the whole attic into every prompt.
Sister agents can inherit decisions, references, checkpoints, source coverage, and task state through bounded session packages and run cards instead of excavating a giant transcript.
Start with public local lite. Add deeper stores when the workload earns them. The agent-facing contract stays still.
Public docs should explain how ContextLattice scales memory: rollups, vectors, durable writes, async recall, retention, compaction, and graph maintenance. Non-public rollout records, deployment secrets, and operating procedures stay out of the public repo.