How it works

Two tools, one loop, on every code write:

• recall: before writing, the agent pulls your standards, ranked by relevance × burn count, trimmed to ~100 tokens.
• capture: when you correct it, the fix is compressed to one line and stored (or its burn count bumps).

Caveman rule format

[TAG] anti-pattern → fix (×N)

Every rule is one terse line. The → fix is mandatory, a rule without a concrete fix is just nagging and gets ignored. The ×N burn count records how many times you have been corrected on it and drives ranking.

Part	Meaning
TAG	One of UI, COPY, CODE, COMMIT, SEC, REQ, PERF
anti-pattern	The habit to avoid
fix	The concrete thing to do instead
×N	Burn count: times you’ve been corrected on this rule. Higher burns rank higher in recall.

Examples:

[CODE]   invented APIs, guessed signatures → verify against the docs first (×4)
[REQ]    gold-plating beyond the ask → build only what's specced; ask first (×3)
[UI]     bespoke UI instead of the design system → reuse tokens + components (×3)
[COPY]   "delve/seamless/robust" LLM slop → plain, concrete language (×2)
[COMMIT] one giant, vague commit → small, conventional: type(scope): msg (×2)
[SEC]    permissive defaults, missing authz → deny by default, least privilege (×1)

What runs where

• Supermemory Local: the shared, on-machine store at http://localhost:6767. Holds the rich memories + local embeddings.
• Ranking is local: self-hosted vector search returns nothing (current release), so remindy lists via documents.list and ranks with a deterministic keyword scorer. Recall needs no LLM.
• Compression: an OpenAI-compatible model at capture time (BYOK). Unreachable? It falls back to a template so capture never blocks.

Two-layer data model

remindy stores a rich memory and injects a caveman projection derived from it.

{ "id": "…", "tag": "COPY", "antiPattern": "…", "fix": "…", "burns": 3, "createdAt": "…" }

• Match & dedup run on the rich memory.
• Only the one-line projection is injected into the agent.
• Projections can be regenerated from rich memories if compression improves.