Recall and capture

Causal is experience memory for AI agents — it records what your agents tried, what happened, and why, then serves that experience into the next decision. This page covers how precedent gets into a decision, and how the result of that decision gets back into the graph.

What does Causal remember?

Most agent memory tools store what the agent knows: preferences, project facts, fragments of earlier conversation. Causal stores what the agent did. It captures decisions, durable context, actions, and observed outcomes from the work your agents do, and connects them to the causes and evidence behind them.

The difference shows up at retrieval. A precedent comes back attached to the outcome it caused, so the agent reads a decision and its consequence rather than re-reading a conversation and inferring one. A path that failed last month arrives labelled as a path that failed.

Causal installs into the harness you already use. See Install Causal in your agent.

How precedent reaches the agent

Two pieces land in every supported harness, and they do different jobs.

The MCP server is the connection. It is how the agent reaches the experience graph to read precedent and write outcomes.

The always-on rule is the trigger. Each harness has its own name for it — a rule, a project rule, a hook — and npx causal init detects the harness you are in and writes whichever one that harness expects. Without it, recall is something the agent has to remember to ask for, and it arrives on a tool call, after the agent has already chosen a direction. With it, the relevant precedent is in context when the agent starts deciding.

A connection without a trigger is a tool the agent may never open.

What the rule does

It recalls relevant experience before acting and captures meaningful outcomes afterward. Three behaviors follow from that placement.

• Small, time-anchored payloads. The agent gets a briefing in a few dense lines rather than a replayed 45k-token transcript. Every session makes the next one cheaper.
• Session-start briefings served from the graph. The agent opens knowing what you tried, what worked, and what broke. Monday picks up where Friday left off.
• Failed attempts returned before the retry. The prior attempt arrives in context with its outcome attached, so a path that already failed is visible before the retry begins. Correct once; the lesson sticks across agents and teammates.

Causal serves precedent and evidence. Your agent still reasons about what to do.

Inside a session-start briefing

The smallest relevant set of precedents, not the session that produced them.

As an illustration, a briefing entry from an earlier billing session would carry where it came from, what happened, what is still outstanding, and what to do next:

That example is written to show the shape. It is not a real record.

The retrieval rules — how relevance is scored, how many precedents come back, and how the payload is bounded — are not yet published.

What gets written back

Decisions, durable context, actions, and observed outcomes. Capture is the write half of the same rule: the work is written to the graph as it goes, so a session that ends — finished, interrupted, or cut off at a usage limit — leaves its experience behind rather than its transcript.

Because the work is recorded as it happens, later changes stay traceable. A session that revises an earlier decision is written into the same graph as the decision it revises, so the history reads as a sequence of changes with their causes. Every entry is signed and traces back to the session that produced it. AI agent governance and audit trail covers what can be rebuilt from that record.

What is written is portable across harnesses. Harness setup covers serving one agent's experience into another, and Agent memory with integrations covers experience arriving from the systems around the agent — Slack, GitHub, Jira, Linear, Notion, Salesforce, and SAP.

Not yet published

The rule itself. What it literally says, where each harness writes it, and how to edit or verify it by hand are not documented yet, and there is no API for writing a custom hook. The scoring behind recall and the payload budget are likewise unpublished.

The commands and the reported install lines on Harness setup are what is settled today.