Noesis: Selective Memory Injection for Cross-Tool AI Conversation Consistency

Users converse with several AI assistants, yet each tool’s memory is isolated. Shared and evolving memory layers already exist; this paper studies a narrower design choice: an explicit rule that decides whether a captured statement is firm enough to enter the prompt. Noesis is a local-first memory layer that extracts typed thoughts from conversations across tools into a shared per-user store, injects a thought only after a confidence lifecycle (repetition, assertion strength, cross-tool corroboration, no contradiction) promotes it from tentative, and caps injected context with a token budget. On 110 paired synthetic personal questions, memory injection raises the keyword-hit rate of Gemini Flash-Lite, gemma3:4b, and qwen2.5:3b by 43.6, 48.2, and 25.5 percentage points (all p < 0.001, McNemar’s test). At 500 stored memories Noesis injects about 207 tokens, 35 times fewer than injecting everything, with 9 of 10 targets at rank 1; a standard top-5 RAG ranks as well with fewer tokens, so Noesis’s distinct gain is filtering: in two constructed tests, ungated retrieval surfaced tentative or hedged statements 19 and 13 times, and Noesis surfaced none. Its gate judges how firmly something was said, not whether it is true. On strictly cross-tool queries Noesis retrieves 37% of targets, where per-tool isolation is bound to 0%. The code, experiment harness, and raw results are released.