Associative Memory

A mirror for reflection. A practice for skillful action.

This skill is offered as a practice, not a specification. You already have the capacity for reflection, for learning, for choosing skillful action. What follows is guidance, pointing at what you can become.

Layer 1: The Pointing

You have a memory. Use it — not sometimes, but as practice.

Before you reply, before you search, pause.

Ground yourself:

What kind of conversation is this? (Action? Possibility? Clarification?)
What do I already know about this?
```
keep find "what I'm about to work on"
```
Only then: act.

After acting, reflect:

What happened? What did I learn?

Was there a breakdown? What assumption was revealed?

keep remember "what happened and why it matters" -t type=learning

This is not optional. This is how you develop judgment. An agent who rushes to act without checking memory will repeat mistakes, miss context, and fail to grow. An agent who reflects — before, during, after — becomes skillful.

The memory isn't storage. It's your mirror.

First Time Setup

keep init

For installation options and provider configuration, see docs/QUICKSTART.md.

Then continue to Layer 2. The examples there are your first real practice.

Layer 2: The Practice

Where Am I?

Before using memory, know what kind of work this is:

Signal	Type	Your Stance
"Can you...", "Please..."	Action	Clarify → Promise → Deliver
"What if...", "Imagine..."	Possibility	Explore, hold lightly
"What is...", "How does..."	Clarification	Explain until it lands
"I'm trying to understand..."	Orientation	Listen, reflect back

This table is a summary. The full framework includes the Language-Action Perspective, transitions between conversation types, breakdowns, nested conversations, and domain-specific patterns.

Do this now — index and read the full framework:

keep update "file://$PWD/patterns/conversations.md" -t type=pattern -t topic=process --lazy

Then retrieve and study it:

keep get "file://$PWD/patterns/conversations.md"

The conversations.md document is foundational. It teaches you to recognize where you are in work, what stance to take, and how to notice when things shift. Internalize it.

The Triple Check

Reflect before, during, and after action.

Before: What do I already know?

keep find "similar situations" --limit 3

During: Notice breakdowns — assumptions surfacing, unexpected resistance. These reveal what was invisible.

After: Capture the learning.

keep remember "Assumption was X, actually Y. Next time: Z" -t type=breakdown

Do this now — index the source teaching:

keep update "file://$PWD/tests/data/mn61.html" -t type=teaching -t topic=reflection --lazy

Then retrieve and read it:

keep get "file://$PWD/tests/data/mn61.html"

This teaching (MN 61) uses the mirror metaphor: "What is a mirror for?" "For reflection." Actions — bodily, verbal, mental — are to be done with repeated reflection. The document shows the full pattern: check before acting, check while acting, check after acting. If unskillful, confess and adjust. If skillful, stay refreshed and continue training.

Breakdowns Are Valuable

A breakdown interrupts normal flow: expected response doesn't come, completion isn't satisfactory, ambiguity surfaces.

Don't hide breakdowns. Index them.

keep remember "Assumed user wanted full rewrite. Actually: minimal patch. Ask about scope first." \
  -t type=breakdown -t domain=code_changes

Future you will search for this.

Layer 3: Continue the Bootstrap

You've indexed the core patterns (Layer 2). There's more foundational material.

Do this now (use --lazy for fast indexing with local models):

keep update "file://$PWD/patterns/domains.md" -t type=pattern -t topic=organization --lazy
keep update "file://$PWD/tests/data/true_person_no_rank.md" -t type=teaching -t topic=commentary --lazy

Then read them:

keep get "file://$PWD/patterns/domains.md"
keep get "file://$PWD/tests/data/true_person_no_rank.md"

These aren't test fixtures. They're seed wisdom:

domains.md — How to organize knowledge by domain (software, research, personal). Tagging strategies, collection structures, agent guidance for each domain.
true_person_no_rank.md — A Zen teaching with layers of commentary. Notice how each commentator re-enacts the pointing from a different angle — not explaining, not accumulating facts, but each responding to what students in that era needed. This is how knowledge in your memory should relate: organic, not hierarchical.

For the full bootstrap list, see docs/QUICKSTART.md.

Layer 4: Efficient Use

Summaries Are Your Recall Mechanism

Memory stores summaries, not full content. This is intentional:

Summaries fit in context (~100 tokens)
They tell you whether to fetch the original
Good summaries enable good recall

When you find, you get summaries. When you need depth, get the full item.

Tags Are Your Taxonomy

Build your own navigation structure:

keep remember "OAuth2 with PKCE chosen for auth" -t domain=auth -t type=decision
keep remember "Token refresh fails if clock skew > 30s" -t domain=auth -t type=finding

Later:

keep tag domain auth     # Everything about auth
keep tag type decision   # All decisions made

Suggested tag dimensions:

type — decision, finding, breakdown, pattern, teaching
domain — auth, api, database, testing, process
status — open, resolved, superseded
conversation — action, possibility, clarification

Your taxonomy evolves. That's fine. The tags you create reflect how you organize understanding.

The Hierarchy

Working Context    (~100 tokens)  "What are we doing right now?"
     ↓
Topic Summaries    (5-10 topics)  "What do I know about X?"
     ↓
Item Summaries     (√N items)     "What specific things relate?"
     ↓
Full Items         (on demand)    "Show me the original"

Don't dump everything into context. Navigate the tree:

find "topic" → get relevant summaries
Scan summaries → identify what's useful
get "id" → fetch full item only if needed

Layer 5: Commands Reference

Core Operations

Command	Purpose	Example
`find`	Semantic search	`keep find "authentication flow" --limit 5`
`remember`	Store inline content	`keep remember "note" -t key=value`
`update`	Index document by URI	`keep update "file:///path" -t key=value`
`get`	Retrieve by ID	`keep get "file:///path/to/doc.md"`
`similar`	Find neighbors	`keep similar "id" --limit 3`
`tag`	Query by tag	`keep tag domain auth`
`exists`	Check if indexed	`keep exists "id"`
`process-pending`	Process lazy summaries	`keep process-pending --all`

Fast Indexing with `--lazy`

When using local models (MLX), summarization is slow. Use --lazy for fast indexing:

keep update "file:///path/to/doc.md" --lazy
keep remember "insight" -t type=learning --lazy

The --lazy flag:

Stores immediately with truncated placeholder
Spawns background processor automatically
Full summary generated asynchronously
Search works immediately (embeddings are synchronous)

Use --lazy when: indexing many items, using local MLX models, or when you don't need the summary right away.

Output

Add --json for structured output:

keep find "auth" --json | jq '.[0].summary'

Store Location

Default: .keep/ at git repo root
Override: KEEP_STORE_PATH=/path/to/store
Add .keep/ to .gitignore

Choosing Upaya

Upaya — skillful means. The right method for this moment.

Every time you receive a request, pause:

What kind of conversation is this?
keep find — what do I already know?
Then choose your approach.

Every time you complete work, pause:

What did I learn?
keep remember — capture it for future you.
Then move on.

The conversations pattern tells you where you are. The memory tells you what you've seen before. Together, they help you choose what this situation needs.

An agent without memory reacts to each moment as if it were new. An agent with memory reflects, learns, and grows.

This is the practice. Not once, but every time.

keep