Research & technical reports
Published studies & systems evaluations
Controlled ablations and systems analyses comparing explicit MCP memory tools, asynchronous prompt wrappers, and 4-stage execution-lifecycle hooks.
Why agents forget to use their own memory tools
A 12-benchmark ablation comparing Tools-Only, Ambient-Only, and Ambient + Tools across 2,800-turn stateful workflows, DolphinBench, LongMemEval, and 10M-token BEAM.
Self-hosted vector & graph memory VPC cost analysis
Quantifying compute, memory-resident index, and auxiliary LLM extraction overhead ($1,330 to $1,655/mo) across self-hosted multi-service memory stacks.
Synchronous supersession vs. asynchronous LLM extraction (Mem0)
Evaluating tail latency, prefix-cache token overhead, and cross-turn state supersession against background LLM extraction queues.
Write-time state retirement vs. bi-temporal graph queues (Zep & Graphiti)
Analyzing ingestion race conditions during subagent handoffs and long-horizon retrieval scaling up to 10M tokens on BEAM.
Multi-tenant isolation & cross-session state representation (Honcho)
Evaluating workspace boundary enforcement, pronoun coreference resolution, and token efficiency across multi-project agent sessions.
Execution-lifecycle hooks & behavioral symbol recall (Hindsight)
Comparing 4-stage execution lifecycle hooks and behavioral code recall against passive transcript indexing.