From 55d6cd53333331b6e58c64c81ce0695828945116 Mon Sep 17 00:00:00 2001 From: artale Date: Thu, 9 Jul 2026 06:48:08 +0200 Subject: [PATCH] docs(tac): add Memory vs Taste architecture pattern Folds the Memory vs Taste concept into loop_engineering.md. Memory is raw storage/retrieval (vector DB, full histories). Taste is higher-order judgment (significance extraction, preferences, lessons). Applied to TAC: skill_health.py as the bridge, deploy policy as taste, 25 containers + qdrant as memory. --- plans/meta-prompts/loop_engineering.md | 13 +++++++++++++ 1 file changed, 13 insertions(+) diff --git a/plans/meta-prompts/loop_engineering.md b/plans/meta-prompts/loop_engineering.md index d3c5187..650ba4f 100644 --- a/plans/meta-prompts/loop_engineering.md +++ b/plans/meta-prompts/loop_engineering.md @@ -338,3 +338,16 @@ Reference: github.com/UnpaidAttention/fable5-methodology — Claude Fable 5 docu - The "write methodology so a weaker model can execute it" pattern is what TAC's loop_engineering.md does for its agents — documented procedures with enforceable gates. **Sleep-time compute reference:** Letta Sleep-time Compute (arxiv 2504.13171) formalizes what TAC's RSI loop already does — offline computation between conversational turns for memory consolidation, reflection, and insight generation. The factory's skill_health.py cron that runs every 6 hours is a primitive implementation. Advanced sleep-time compute would add structured reflection, cross-skill pattern discovery, and automated methodology refinement. + +## Memory vs Taste — two-layer agent architecture + +Memory is raw storage and retrieval — full conversation histories, documents, events, vector databases (e.g. qdrant:6333, GBrain). It scales to tens of thousands of files but risks overwhelming the model with noise. + +Taste is the higher-order judgment layer — selectively understanding *why* something matters, extracting significance, preferences, and lessons rather than archiving everything. It enables better decision-making and more human-like reasoning. + +In practice: Hermes, OpenClaw, and Claude Code combine scalable memory (graph-based synthesis via GBrain or vector DB) with taste mechanisms to prioritize relevant insights, reducing hallucination and improving output quality over brute-force context expansion. + +**Applied to TAC's factory:** +- **Memory layer:** 25 containers, 63 skills in three hermes agents, full RSI skill-health pass/fail history, qdrant vector DB, forgejo issue tracker, complete conversation logs — everything stored. +- **Taste layer:** The signed-action deploy policy (gating what `:8098` can accept), the `auto_patch_proven` threshold (3 receipts × 2 skill types), the verifier/reviewer gates in the ADW pipeline, and every correction that marks stale entries as SUPERSEDED. These are taste — judgment applied to raw data. +- The RSI loop bridges both: skill_health.py reads memory (pass rates), applies taste (below 80% = degraded), triggers action (repair). The taste layer is where TAC's value lives — without it, the factory is just an expensive archive.