Restore Project Context with Semantic Memory

Context Restoration is a skill for intelligent, semantic-aware context retrieval and reconstruction across complex, multi-agent AI workflows, aimed at preserving and reconstructing project knowledge with high fidelity across long-running projects. It takes a context source (vector database or file system), a project identifier, a restoration mode (full, incremental, or diff to compare and merge context versions), a token budget (default 8192), and a semantic-similarity relevance threshold (default 0.75).

Its retrieval strategy combines semantic vector search (multi-dimensional embeddings across text, code, and architectural diagrams, ranked by cosine similarity), multi-signal relevance ranking (semantic similarity, temporal decay, and historical decision impact combined into a composite score), and token-budget-aware rehydration that prioritizes context components - project overview, architectural decisions, technology stack, recent agent work, and known issues - and loads them in priority order until the budget is spent. It also covers session-state reconstruction (agent workflow state, decision trails, multi-agent collaboration history), context merging with three-way conflict resolution and provenance tracking, incremental/lazy loading and streaming for large projects, integrity checks (cryptographic signatures, semantic consistency, version compatibility), and performance optimizations like caching and efficient vector-search indexing.

Two reference workflows illustrate its use: project resumption (retrieve the most recent context, validate it against the current codebase, selectively restore relevant components, generate a resumption summary) and cross-project knowledge transfer (extract semantic vectors from a source project, map and transfer relevant knowledge, adapt it to the target project's domain, and validate that the transfer actually applies).

Use it when resuming a long-running, multi-agent project that needs its prior context - decisions, architecture, recent work - reconstructed with fidelity rather than re-derived from scratch, or when transferring relevant knowledge from one project to a related one. It is not a substitute for validating restored context against the actual current codebase - the project-resumption workflow explicitly includes that validation step rather than trusting stored context blindly.

Input is a project identifier, restoration mode, and optionally a semantic query. Output is a reconstructed context object containing the prioritized, budget-fitted components, or a resumption/knowledge-transfer summary. It's invoked like this:

### Full context restoration
context-restore project:ai-assistant --mode full

### Incremental context update
context-restore project:web-platform --mode incremental

### Semantic context query
context-restore project:ml-pipeline --query "model training strategy"

It's designed to plug into RAG (Retrieval Augmented Generation) pipelines, multi-agent workflow coordination systems, continuous learning systems, and enterprise knowledge management setups.

Teams running long-running, multi-agent AI projects who need reliable context restoration between sessions or knowledge transfer between related projects, rather than re-deriving project context from scratch each time.