Procedural Memory

Procedural memory is the how-to layer: the runbooks, checklists, and skills the agent reuses when the same task comes back. In human terms, it's the difference between knowing that a bike has two wheels and knowing how to ride one.

Facts vs. procedures

For the issue-#9 example, you don't want a loose rule like "if the issue is open and assigned, implement." You want the real procedure, with all its checks:

procedure: approval-gated-implement
  1. fetch the issue and its comment thread
  2. confirm a fix proposal exists
  3. confirm the approval comment appears AFTER the proposal
  4. read the getting-started doc
  5. work only inside imports/api/ and imports/ui/
  6. run the focused tests
  7. open a draft PR, comment back with the result

Compare the two kinds of knowledge directly: "implementation requires explicit approval" is semantic memory — a fact. The gated sequence above is procedural — a way of working.

The least obvious memory type

People rarely call this layer memory at all. They call it tools, skills, automation. But it is stored knowledge about how work should happen, and it lives in more places than you'd think:

• Skill files — the most recognizable form: named, reusable procedures loaded when the task matches
• Tool schemas — a schema that requires an issue number and an approval-comment ID forces the procedure structurally
• Orchestration code — workflow graphs, agent loops, and approval gates encode procedure even though the agent never sees it as prose

In the wild

This layer is the memory-stack view of the harness's skills primitive — the same artifact seen from two frames. Harness engineering asks "how do we make work repeatable?"; memory architecture asks "how does know-how survive across sessions?" The answer to both is the same files.

The pieces are all here — separately

Working memory for what's active now. Episodic for what happened. Semantic for what's true. Procedural for how work is done. Four stores, four jobs.

But durable stores are just a filing cabinet — and the model never reads the cabinet. It only ever works at the desk. Everything now depends on what gets pulled out and set down in front of it.