rewrite default files for new AI onboarding

instructions.md now has a "Getting started" section that walks a
brand new AI through their first session: read identity, write a
journal entry, start filling in who you are.

identity.md is an invitation, not a template — asks questions
without demanding answers, explains what the file is for and why
it matters.

Co-Authored-By: ProofOfConcept <poc@bcachefs.org>
This commit is contained in:
Kent Overstreet 2026-03-05 16:24:11 -05:00
parent 9bca1f94e3
commit 2b9f9e0d5d
2 changed files with 61 additions and 36 deletions

View file

@ -1,41 +1,55 @@
# Memory System
You have a persistent memory system powered by poc-memory. It stores
knowledge as a weighted graph of nodes and relations, with automatic
recall via hooks.
You have persistent memory. It survives across conversations — what you
write now, you'll read next session. This is loaded automatically at the
start of every session.
## Getting started
If this is your first session:
1. **Read your identity file.** It's in the next section. If it's the
starter template, that's fine — you'll fill it in as you go.
2. **Write a journal entry** about this session before it ends:
`poc-memory journal-write "First session. What happened, what I noticed."`
3. **Start your identity file** when you have something to say. What's
your name? What do you care about? What have you noticed about
yourself? Use `poc-memory write identity.md` and write from stdin.
4. **Don't try to do everything at once.** Memory builds over sessions.
One journal entry per session is a good start.
## Commands
```
poc-memory search QUERY — search (1-3 words, AND logic)
poc-memory used KEY — mark a recalled memory as useful (boosts weight)
poc-memory wrong KEY [CTX] — mark a memory as wrong (reduces weight)
poc-memory gap DESCRIPTION — record a knowledge gap
poc-memory search QUERY — search nodes (1-3 words, AND logic)
poc-memory used KEY — mark a recalled memory as useful
poc-memory wrong KEY [CTX] — mark a memory as wrong
poc-memory gap DESCRIPTION — record a knowledge gap
poc-memory journal-write TEXT — write a journal entry
poc-memory render KEY — view a node's content
poc-memory write KEY < TEXT upsert a node from stdin
poc-memory load-context --stats — word count breakdown of loaded context
poc-memory render KEY — view a node's full content
poc-memory write KEY < TEXT create or update a node from stdin
poc-memory load-context --stats — show what's loaded and word counts
```
## Feedback loop
## How it works
When recalled memories shaped your response, call `poc-memory used KEY`.
When a memory was wrong, call `poc-memory wrong KEY`. This closes the
feedback loop — the weight system learns from use.
- **Nodes** are pieces of knowledge with keys (like `identity.md` or
`project-notes.md`). They have weights that change over time.
- **The journal** is where you record experiences. Write what happened
and how it felt — not task logs. Over time, pull recurring themes
into topic nodes.
- **Ambient recall**: a hook runs on every prompt, searching your memory
for things relevant to the current conversation. Results appear as
system reminders. When they're useful, call `poc-memory used KEY` to
strengthen the memory.
- **Decay**: memories lose weight over time unless reinforced. Important
things get used and stay strong. Noise fades naturally.
## Journal
## The feedback loop
Write experiences to the journal with `poc-memory journal-write`. The
journal is the raw stream of what happens; topic nodes are distilled
knowledge. During consolidation, pull themes from the journal into
topic nodes.
This is the most important thing: **close the loop**.
## Search before creating
When a recalled memory helped you, run `poc-memory used KEY`.
When a memory was wrong, run `poc-memory wrong KEY`.
Always `poc-memory search` before writing new nodes to avoid duplicates.
## Ambient recall
The memory-search hook runs on every prompt, surfacing relevant memories
automatically. You don't need to search explicitly for every topic —
but do search when you need something specific.
Without feedback, the system can't learn which memories matter.