Back to AI Research

AI Research

Blast Radius | AI Research

Key Takeaways

  • Blast Radius is a memory management layer designed to make agentic coding more affordable and sustainable by reducing the number of tokens consumed during lo...
  • Agentic coding faces growing problems of affordability and wasted tokens.
  • We introduce Blast Radius, a predictive memory management layer that estimates an incoming prompt's reach through coupled context and code channels.
  • NECROPHORESIS enables reversible eviction by archiving dead context verbatim, while Recurring Dead Matter (RDM) identifies and buries repeatedly occurring transcripts.
  • We formulate reversible context eviction over a Polish context space, providing a measurable foundation for retention, recurrence, and eviction while connecting context entropy to resurrection probability.
Paper AbstractExpand

Agentic coding faces growing problems of affordability and wasted tokens. We introduce Blast Radius, a predictive memory management layer that estimates an incoming prompt's reach through coupled context and code channels. NECROPHORESIS enables reversible eviction by archiving dead context verbatim, while Recurring Dead Matter (RDM) identifies and buries repeatedly occurring transcripts. We formulate reversible context eviction over a Polish context space, providing a measurable foundation for retention, recurrence, and eviction while connecting context entropy to resurrection probability. Across seven OpenAI models, Blast Radius reduced token consumption by 17-26%, achieved the lowest overflow rate among tested policies, and remained byte exact reversible. Of 450 buried bodies, 378 were recurring dead matter and zero were recalled. Blast Radius operates beneath HCRC, determining which records to bury and how far an incoming prompt may reach into the codebase. This work contributes to the broader goal of Algosophy: making large language models and agentic coding more reusable and sustainable.

Blast Radius is a memory management layer designed to make agentic coding more affordable and sustainable by reducing the number of tokens consumed during long-running coding tasks. Developed by MY Pitsane and Hope Mogale, the system predicts how much context an incoming prompt will require and selectively archives "dead" information to free up space in the model's context window.

Managing Context Through Prediction

Agentic coding loops often re-submit entire conversation histories, including tool outputs and file contents that are no longer relevant to the current task. This leads to high token costs and wasted bandwidth. Blast Radius addresses this by estimating the "blast radius" of a prompt—a two-channel calculation that determines both the temporal reach (how many context tokens are needed) and the structural reach (which parts of the codebase are affected by code edits). By calculating these needs before a turn runs, the system identifies which information is truly necessary and which can be safely moved out of the active context window.

Reversible Forgetting with Necrophoresis

Unlike traditional memory management techniques that permanently delete or summarize information, Blast Radius uses a process called Necrophoresis to ensure that any evicted data remains recoverable. When the system identifies "dead" context—information that is unlikely to be needed again—it archives the data verbatim into an on-device storage area called a "midden."
The system replaces the original content with a compact "scent skeleton," which acts as a placeholder. If the model later requires the original information, it can perform an "exhumation" to restore the data exactly as it was. Because this process is byte-exact and lossless, the risk of losing critical information is bounded by a small, fixed cost for exhumation.

Handling Recurring Dead Matter

Beyond general context, the researchers identified that coding loops frequently generate repetitive transcripts, such as repeated test suite outputs or build logs. These "Recurring Dead Matter" (RDM) instances are identified by the system and buried using Laplace’s rule of succession. By tracking how often these classes of transcripts are exhumed, the system learns to automatically archive older instances of routine tasks. This prevents the context window from becoming cluttered with redundant information that does not contribute to the model's current objective.

Performance and Results

In tests conducted across seven OpenAI models, Blast Radius reduced token consumption by 17–26%. The researchers reported that the system achieved the lowest overflow rate among the policies tested. Out of 450 "buried" items during their evaluation, 378 were identified as recurring dead matter, and zero items were recalled, suggesting that the system effectively distinguished between transient noise and necessary information. The researchers note that this work is part of a broader effort in "Algosophy," which aims to improve the reusability and sustainability of large language models in software engineering.

Comments (0)

No comments yet

Be the first to share your thoughts!