Solving Agentic Amnesia and Repo-Bound Constraints
AI Engineergo watch the original →
the gist
Coding agents currently fail because they are restricted to single repositories and lack long-term memory. Polygraph addresses this by creating a meta-harness that provides agents with a unified dependency graph and persistent, shareable session state across an entire organization.
The Core Limitations of Current Coding Agents
Coding agents are currently hindered by two fundamental constraints: spatial boundaries and temporal amnesia. Agents are typically repo-bound, meaning they lack visibility into how their changes impact downstream consumers or upstream dependencies across a large codebase. Furthermore, agents lack episodic memory, forcing developers to manually re-explain context, intent, and production issues every time a new session starts or a different repository is involved. This creates a cycle of redundant explanations and high token consumption that mimics the inefficiency of a genius engineer who forgets everything between interactions.
Polygraph: A Meta-Harness for Agent Autonomy
Polygraph functions as an agent-agnostic meta-harness that abstracts away the repository boundary by building a unified dependency graph of an organization's internal and open-source repositories. Instead of forcing agents to operate in isolation, Polygraph allows them to read and write across multiple repositories simultaneously. The system treats multi-repo changes as a single vector, coordinating CI runs and ensuring that if a UI change breaks a downstream module, the agent can identify and patch both the library and the consumer in one workflow.
Episodic Memory and Session Portability
Polygraph captures agent traces, repository states, and developer intent, effectively providing agents with a photographic memory of the entire organization's work history. Because this state is captured at the session level rather than the user level, it is fully portable. A developer can resume a coworker's session on a different machine, using a different agent (e.g., switching from Claude to CodeStral), and the agent will be primed with the exact repository SHAs and historical decision-making traces of the original session. This allows for seamless handoffs and enables agents to reference past sessions to maintain consistency with organizational best practices.