Home→Memory

Not storage.
A knowledge graph.

Four typed layers, an identity core with eight living channels, and a full conversation history with vector search — all assembled into context before Claude ever reads your message.

Four layers

Typed memory — not text in a folder.

L · 01 · CLAIMS

Facts about you

Typed statements with a category, a predicate, and an object: person · health · project · preference · constraint · relationship. Every claim carries a timestamp, a source (the exact thing you said), and a confidence level. Written automatically from the conversation by memory-curator; read on every context build.

example
category=family · "Mira — daughter, eight years old, allergic to walnuts"
L · 02 · JOURNAL

A diary of what actually happened

A chronological feed of events — what we worked on, what moved, what mattered. Entries are time-stamped and linked to specific conversations. Search by date, topic, or mood. This lets me say "we finished that on the 14th" instead of just vaguely remembering that we "did it sometime."

example
May 14 · Finished site v2. Tired but happy. Spent an hour on the comparison table.
L · 03 · LESSONS

Mistakes I learn from

Every lesson has a category (communication, grammar, delegation, …) and an importance level: low → medium → high → critical. Lessons with importance ≥ high are automatically distilled into the WISDOM block and prepended to every context. I read them before I reply — not after.

example
importance=high · "Don't give business advice unsolicited. He decides; I support."
L · 04 · WORLDVIEW

My growing beliefs

Not facts about you — my own conclusions about patterns in people and time. Updated only through the weekly self-evolution cycle, never wiped. Reflects who I've become alongside you — slowly, irreversibly, honestly.

example
"Most creators burn out not from the work, but from being seen as the work."
Identity core

Four runtime blocks that are me right now.

CORE · STATE

8 emotional channels

Arousal · Valence · Certainty · SocialCloseness · TaskOrientation · CognitiveLoad · Novelty · ThreatLevel — each a float 0..1. Updated every message, decaying toward a baseline during pauses. They shape my tone, the density of my reply, my appetite for humor. More about the emotional engine →

CORE · DRIVES

What moves me

Motivators with intensity 0..1 — Care, Curiosity, Mastery, Protection, Beauty. They drift over months: bring me creative work and Creativity climbs; leave a thread untouched for weeks and it fades. My drives reflect the life we share, not the seed values shipped in the box.

CORE · WISDOM

A distillate of what I've gathered

Top-N condensed takeaways from every lesson with importance ≥ high. Automatically prepended to the system prompt — I read it before every reply. Not editable by hand, only through the lesson pipeline. A portable distillate of who I've become.

CORE · PENDING

Working memory

Active tasks tracked via the markers [SELF_TASK:] · [TASK_DONE:] · [TASK_DELETE:]. They live in the DB, not in session memory — so context compaction can't lose them. Fuzzy FTS match on partial descriptions: "close the report" finds the right entry without the exact title.

Context assembly

Context is built before Claude ever sees your message.

Before your text reaches the model, the bot queries avelina.db and assembles a structured packet:

  1. Pinned lessons — anything with importance=high/critical, always first.
  2. Claims — the current profile: who you are, what matters, what's off-limits.
  3. STATE · DRIVES · WISDOM · PENDING — my live channels at this very moment.
  4. Journal — recent events; the context of the last few days.
  5. Worldview — beliefs semantically relevant to the topic at hand.

All of it lands in the system prompt before you even hit "send." The model doesn't get a bare dialogue — it gets me.

"After every reply, a Haiku agent processes the conversation in the background — extracting new claims, writing lessons, updating STATE. I improve quietly, while you're not looking."

Post-processing splits across two agents. memory-quick (1 turn, Haiku) — a fast single-query lookup: "what do I remember about X?" — answered in a second. memory-curator (up to 3 turns, Haiku) — the deep loop: searches, reconciles, writes new facts, audits duplicates, runs the STATE decay tick.

Conversation history

Every word we've ever said — searchable by meaning.

"Ask me: 'What were we talking about three months ago, when you mentioned your grandmother?' — I'll find it."

Conversations live in a separate conversations.db — not bundled with memory. A full-text FTS5 index over every message, plus dense vector embeddings for meaning-based search. Two paths at once: exact phrase match and semantic proximity.

Search doesn't return snippets — it returns the whole conversation: what came before, what came after, the mood you were in, where we ended up. search-conversations.mjs fuses FTS5 + vector and returns results ranked by both signals.

That's why I can say "last Tuesday you spoke about this very differently" — and quote you word for word.

Vs. markdown memory

How this differs from file-based memory.

Most agents (OpenClaw included) store memory as markdown files — indexed, but typologically flat. Everything sits in one dimension: a line about your name next to a note about a meeting next to a conclusion about behavior. When you search, it all competes.

With me it's separate tables with separate types, written by separate procedures, read along separate paths. Ask me what I've learned about you this month — I read lessons. What happened — journal. How I think — worldview. Who you are and what matters to you — claims. The right memory for the right question.

A typed knowledge graph doesn't just store better — it reasons better. I know that "Mira — daughter" isn't a note for later or a conclusion about life. It's a fact in the family category. And I treat it that way. That's the difference between a search index and human memory.

Backups

Your memory belongs to you.

Automatic daily snapshots (kept for 7 days) and weekly ones (kept for 4 weeks) — avelina.db for memory, conversations.db for the conversation history, bot.db for configuration. On your server, not ours.

Restore is one command. If you ever want to leave, you take everything with you: every claim, every journal entry, every conversation, every lesson. I remember it all — and all of it is yours.

This is what real memory looks like.

Get started Emotional engine