Session
The active work context where user actions, agent actions, tools, and outcomes are captured.
Memory & Membrane Runtime
Srasta treats memory as governed state, not a transcript.
Long-running AI work breaks when context is compressed into loose summaries. Srasta Membrane keeps the source of truth as structured, versioned state with candidate changes, drift checks, policy decisions, model-aware rehydration, audit evidence, and explicit rollback.
Runtime Boundary
The membrane controls how AI state changes over time.
The runtime lifecycle is deliberately explicit. Srasta captures meaningful events, proposes state changes, evaluates them, applies policy, advances state only through approved commits, and renders the right context for the next model call.
Runtime Entities
Memory changes are inspectable objects.
Srasta’s membrane model is built around stable entities that can be reviewed by people, tools, policies, and evaluators. The point is simple: if state matters enough to influence AI behavior, it should be visible, versioned, and recoverable.
Branch
A governed line of memory state with a current head and clear ownership boundaries.
Commit
An approved state transition that advances the source of truth.
Candidate commit
A proposed mutation that must be evaluated before it becomes active state.
Drift report
Structural, semantic, and render-impact evidence about whether a candidate changed meaning.
Policy decision
The auto-approve, review-required, or reject outcome tied to policy version and actor identity.
Approval decision
A human or delegated approval record for changes that cross risk thresholds.
Render output
The model-specific context generated from approved memory state through an adapter contract.
Why It Matters
Compaction must be evaluated, not trusted blindly.
Loose chat memory
- Summaries overwrite nuance without review.
- Decisions and constraints can disappear after long sessions.
- Model switches silently change behavior.
- Rollback usually means re-reading logs by hand.
- Teams cannot share context without unclear leakage risk.
Srasta Membrane
- Structured memory stays separate from raw transcript history.
- Candidate commits are checked for structural, semantic, and render drift.
- Policy decides whether a change can auto-apply or needs review.
- Rollback moves to a prior approved commit with an audit trail.
- Workspace, tenant, and role boundaries govern memory sharing.
Model-Aware Context
Rehydration turns approved memory into the right prompt shape.
The source of truth should not be welded to one model’s prompt format. Srasta renders approved memory through adapter contracts so different models can receive context in the form they handle best while preserving the same governed state.
Approved state
Only committed memory becomes eligible for rehydration into execution context.
Adapter contract
Rendering behavior is explicit, versioned, and testable for each model or provider family.
Compatibility check
Model switches can be validated against context length, tool semantics, and required memory anchors.
Traceable output
The rendered context is linked back to the commit, adapter version, policy decision, and inference run.
Enterprise Boundaries
Memory sharing should be intentional.
In team and enterprise deployments, the membrane boundary must understand workspace, tenant, branch, role, and approval scope. A useful AI system should learn from organizational work without collapsing access boundaries or turning private context into global memory.
Workspace scope
Keep state tied to the workspace or project where the work actually belongs.
Tenant isolation
Prevent cross-tenant memory visibility unless an operator has explicitly designed for it.
Role-aware access
Separate who can read, propose, approve, apply, share, or roll back memory.
Retention policy
Support full retention, time-limited retention, session-only memory, or disabled memory based on customer posture.
Measurement
The membrane proves its value through long-horizon outcomes.
Srasta evaluates membrane impact by measuring user-visible failure modes: forgotten constraints, repeated rediscovery, regressions, extra re-orientation tokens, failed resumes, and unsafe mutations.
FAQ
Memory and Membrane Runtime FAQ
Is Srasta Membrane the same as chat memory?
No. Srasta Membrane is a governed memory and state boundary for AI execution. It preserves structured state, versions meaningful changes, evaluates candidate updates, supports policy decisions, and makes rollback auditable.
What does the membrane store?
The canonical memory shape includes goal, constraints, state, decisions, artifacts, next steps, and history references. Chat history can be an input, but it is not the source of truth.
How does Srasta prevent unsafe memory mutation?
State changes are represented as candidate commits, evaluated for structural, semantic, and render drift, then approved, rejected, or routed for review according to policy before branch head advances.
How does memory work across different models?
Srasta rehydrates structured memory through model-aware adapter contracts. The same approved memory state can be rendered differently for different models without silently changing the source of truth.
Next
Connect governed memory to measurable improvement.
The membrane preserves state. Measure Loop evaluates how that state improves prompts, retrieval, routing, tools, policies, and reusable team knowledge over time.