Abstract
AI agents fail over time for three reasons: their memory fills with stale history, they re-process the same goals in loops, and injected information corrupts what they know. LOCI solves all three. LOCI assigns every memory fact a heat score that decays over time by tier — critical facts survive 30 days, routine facts 3 days, nothing is ever deleted. Facts are placed across hardware tiers: the most important live in GPU memory during inference, the rest wait in RAM or on disk until needed. When an agent's goal changes, the decay rates shift — facts relevant to the current goal are preserved longer, everything else fades naturally. A circuit breaker compresses each new task into a 25-character fingerprint. If that fingerprint has been seen before, the agent stops before inference is called — zero tokens wasted, zero billing, the loop is broken. When the breaker fires, it identifies which process caused the ramp-up, classifies it, and kills it by port. The Imagination Engine combines facts that are semantically far apart but both trusted and important. The formula is: dream_score = (1 - cosine(A,B)) x heat_A x heat_B. High distance between trusted facts produces the highest synthesis value. The output is stored back into memory and seeds the next generation of combinations. No pathway is rejected for impossibility. A security layer adapted from web UI tooling prevents injection: base facts always win over challengers, injected overrides are quarantined, profile drift is detected by SHA-256 seal. This system was built by one person, with one AI collaborator, on consumer hardware, in Dhaka, Bangladesh. It outperformed GPT-4 with no memory by 18.2 percentage points across 500 standard benchmark questions. With a local model and no cloud inference, it reached 96.6%. This disclosure places the full architecture in the public domain.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Sanyal, Arko and Independent Research, "LOCI: A Heat-Decaying, Goal-Conditioned Memory Architecture for Long-Horizon AI Agents — with Imagination Engine, Circuit Breaker for Recursive Rumination Prevention, and the First Documented Case of Sustained Human-AI Co-Invention via Persistent Memory", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/11666