A new chat loses the working position.
The prior conversation may still exist, but the next Agent does not receive the relevant decisions, requirements, and unfinished work automatically.
PRODUCT / CONTINUITY
A memory layer for work that spans conversations.
TMCRA gives Agent systems a controlled place to retain project state, user requirements, prior progress, and source evidence—then recover only what matters to the current question.
01 / WORKFLOW
One continuity thread, from a completed turn to the next action.
- 01CAPTURE
Record the completed user and Agent turn as separate, attributable messages.
- 02SCOPE
Keep stable user facts in Global memory and project work inside its own Project Scope.
- 03RECALL
Use the current prompt to select relevant evidence from the allowed scopes.
- 04CONTINUE
Deliver bounded Prompt Evidence so the Agent resumes from the right project state.
02 / CONTINUITY GAPS
A stored history still needs a way back into the work.
Long-running Agent work breaks in recognizable ways. TMCRA treats these as product constraints, not decorative use cases.
Similar text does not resolve changing state.
A useful memory layer must preserve who said something, when it changed, where it belongs, and which source supports it.
A log is not yet prompt-ready evidence.
The current task needs a bounded selection that separates user requirements from Agent progress and carries both into the next action.
03 / PRODUCT SURFACES
Use memory without operating a memory server.
The same memory model appears at different densities across the desktop app, web console, API, and native adapters.
Continue and manage
See the current project, prior progress, connections, memory, imports, usage, and account state in one application.
View release status →Inspect and govern
Inspect scopes, sessions, evidence, recall results, API Keys, and server-reported quota.
Open console →Build into your runtime
Use the stable HTTP contract or preview lifecycle clients without replacing your model or Agent runtime.
View integrations →Trace the evidence
See where a memory came from, which scope it belongs to, and how the returned evidence was composed.
View architecture →04 / MEMORY APPROACHES
Choose the memory boundary before choosing the interface.
The comparison focuses on continuity, changing state, isolation, and provenance—the parts that decide whether earlier work can safely affect the next answer.
ApproachAcross conversationsChanging stateBoundary and source
Context WindowCurrent conversationPrompt order onlyConversation boundary
Vector RAGRetrieves stored chunksSimilarity-ledDepends on index design
TMCRAGlobal + ProjectSource + Fast + SlowActor · Session · Time · Source
05 / WHERE IT FITS
Four systems that depend on remembered work.
Preferences, commitments, and ongoing projects
Carry stable user information across tools while keeping every project in its own scope.
Plans, actions, and unresolved work
Specialized Agents can share project state while preserving their own actor and session provenance.
Governed memory for repeated workflows
Use controlled scopes, attributable evidence, deletion, export, usage, and audit surfaces.
Experience that survives task boundaries
Preserve events, corrections, and operational context as a system moves between sessions and tasks.
06 / BOUNDARIES
Ten projects do not become one graph.
Global memory carries stable user information. Every project keeps its own scope, and every conversation remains a session inside that project.