Researchers at Mankind Research Labs in South Africa published Blast Radius on 7 August 2026, a predictive memory-management layer that cuts token consumption in agentic coding loops. Tested across seven OpenAI models, it reduced token use by 17–26%, achieved the lowest context-overflow rate among comparison policies, and maintained byte-exact reversibility—every eviction could be undone without information loss.
The core problem is structural. Each turn in an agentic loop re-submits the entire conversation: system prompt, tool schemas, prior file reads, diffs, stack traces. Cost grows with the sum of all prior context. A file dumped in full to fix a typo on turn 3 still occupies 2,000 tokens on turn 40, long after that sub-task closed. Authors MY Pitsane and Hope Mogale call this "pure epistemic entropy"—bandwidth consumed by noise rather than signal. The model pays attention-quadratic and dollar-linear costs to process context that can no longer affect the next action.
Existing strategies trade a recoverable cost (tokens) for an irrecoverable one (information). Sliding-window truncation drops the oldest tokens indiscriminately. Summarization compresses context through a second model call—itself fallible—and cannot be reversed if it discards something that matters later. Blast Radius estimates, before a turn executes, how far that turn will reach, then uses that estimate to license an eviction with bounded downside.
| Strategy | Reversible? | Information Loss | Extra Cost | Mechanism |
|---|---|---|---|---|
| Sliding-window truncation | No | High — oldest tokens dropped indiscriminately | None | Discard oldest context at window boundary |
| Summarization | No | Medium — lossy compression via second model | Extra model call (itself fallible) | Compress prior context through a second model call |
| Blast Radius / NECROPHORESIS | Yes | None — context archived verbatim | O(1)·κ per exhumation (fixed per-token cost) | Predictive eviction; skeleton replaces archived block in-window |
The architecture uses two channels. The context channel predicts the increment to the retained working set, converts that into an eviction budget, and applies a knapsack policy to select which dead context to archive. The code channel computes the churn-weighted set of files and symbols the turn's edits will touch through the dependency graph—the structural impact surface—and can checkpoint before changesets grow too large to review. Both answer: what is causally coupled to what I'm about to do?
Eviction is handled by NECROPHORESIS. Instead of discarding, it archives context verbatim and replaces it in-window with a skeleton. If wrong, restoration costs O(1)·κ—a single exhumation step at fixed per-token cost. Recurring Dead Matter (RDM) handles recurrence: it identifies near-identical transcripts injected repeatedly and buries entire classes using Laplace's rule. In test runs, 378 of 450 buried bodies (84%) were classified as recurring dead matter, and zero were recalled—every eviction held.
| Metric | Result |
|---|---|
| Token reduction vs. baseline | 17–26% |
| Context-overflow rate | Lowest among all comparison policies |
| Bodies buried (total) | 450 |
| Classified as Recurring Dead Matter (RDM) | 378 (84%) |
| Evictions recalled (reversed) | 0 |
| Restoration cost per exhumation | O(1)·κ (fixed per-token) |
Blast Radius operates beneath the HCRC gate, which controls admission. Division of labor is intentional: HCRC handles admission control, Blast Radius handles eviction selection. The information-theoretic framing connects context entropy to resurrection probability, arguing reversible forgetting reduces epistemic entropy without the finality that makes lossy approaches risky.
The work is a progress report toward Algosophy, a research program at Mankind Research Labs for making agentic coding reusable and sustainable. No open-source release is announced. The paper gives token-count reductions only, not dollar costs. Teams will need to map the 17–26% reduction to their own per-token pricing and call volume.
Takeaway: reversible eviction with a two-channel reach estimator is more defensible than truncation or summarization—but the 17–26% figure remains a benchmark to validate against your own loop depth and model mix until the code ships.