Memory

Two files that make corrections persist — a growing registry of patterns organized by category, and a change propagation map that prevents drift. AI stops repeating the same mistakes.

2 files · 10 min setup · shared/learnings.md · shared/dependency_map.md

The problem

Corrections don't persist. You fix the same mistake three times across three sessions. Each correction is local to one conversation. There's no system for capturing what you learned and applying it everywhere.

Most people re-explain their preferences every session. That's not a workflow problem — it's a missing layer. Memory is the structural fix: two files that accumulate corrections over time and make them permanent constraints for every agent.

What gets encoded

Two files working together: one captures what you've learned, the other tracks what needs updating when things change.

learnings.md

A growing registry of corrections organized by category — Voice & Tone, Conceptual Framing, Process & Workflow, Content Quality, Build Patterns, Agent Design. Each entry captures what went wrong, why it happened, the reusable pattern, and when it was caught.

dependency_map.md

A change propagation registry. Maps which files depend on which. When one file changes, you know exactly what else needs updating. Prevents drift between assets, plans, and agent prompts that reference the same source of truth.

Real example

Here's an entry from Agent Prime's actual learnings file — a voice pattern caught during a Writer draft review:

shared/learnings.md — Voice & Tone

### V7: No "not X but Y" constructions
**Source:** Writer draft review, 2026-02-15
**Pattern:** AI defaults to defining things by contrast ("not a tool
but a partner"). This is weak writing. State the positive thing
directly. Say what it IS, not what it isn't.
**Propagation:** All agent prompts, Writer voice rules

Every agent reads learnings.md before producing output. That entry alone eliminated an entire class of weak framing across every artifact — permanently, from that session forward.

How to set it up

Create shared/learnings.md with category headers. Start empty. Every time you correct AI output, capture the pattern — not just the fix, but what class of mistake it was:

## Voice & Tone
## Conceptual Framing
## Process & Workflow
## Content Quality

Create shared/dependency_map.md listing which files read from which. When you change one, update its dependents. The format is simple — source file, what depends on it, what to check after a change.

The rule that makes it compound: when AI corrects, the first action is to capture the pattern, not fix the immediate output. Fix comes second. Miss the capture and the same class of error recurs next session.

What you get

shared/learnings.md

Your accumulated correction library. Every agent reads this before producing anything. Hard constraints, not suggestions.

~327 lines · grows ~5–10 new rules/month

shared/dependency_map.md

Change propagation registry. Consult after modifying any file. Prevents drift across the system.

~60 lines · stable