Six months of agent-memory provenance fields nobody could see: a Reddit stranger found the gap
Neither-Witness-6010 · reddit · 2026-10-19
The author built CogniCore, a memory layer for coding agents that makes provenance first-class: every entry carries sourceagent/sourcetask, timestamps, confidence, supersedes/invalidatedby links, and a state enum separating 'asserted' from 'independently verified' outcomes.
The lesson: a field is not a feature until the interface surfaces it. After six months of infrastructure, their recall tool returned only outcome/relevance/action/error — none of the provenance — and a stranger on Reddit spotted the gap before they did.
Open design question: how much provenance should travel with a recalled memory? Their current stance: always include source + verification state, attach supersession links only when actually superseded.
Honest limits: retrieval is in-process TF-IDF/BM25 (no vector DB); on LongMemEval multihop, strict recall@5 goes 78.8%→85.2% at 5 chunks and 87.2%→92.8% at 10, with the baseline catching up at 20 — reported rather than hidden. MIT-licensed at github.com/cognicore-dev/cognicore-env
More from coding & agent
- vLLM 0.32.0 adds structured-decision API reading first-token probabilities — AccBalanced · 2026-10-19
- 10 open-source GitHub repos for fully local speech-to-text, led by Whisper at 110K stars — victor_explore · 2026-10-19
- All Things Open talk: hardening agent security with Docker sandboxes — HowDevelop · 2026-10-19
- Nace AI launches Drex Max, a screenshot-in decision model for agents at 174 ms and $0.06/M tokens — testingcatalog · 2026-10-19
- Watch Claude Opus 5.5 build a game-ready bush via Blender MCP in 30 seconds — sidahuj · 2026-10-19
- Anthropic: Engineers' New Role Is Saying "No" to Claude — But Who Trains the Next Generation? — sqcai · 2026-10-19