Operation Scarecrow
Everything under AI memory and Company Brains, organized by the four kinds of memory an org actually has, plus the case studies that don't sort neatly into just one.
The Argument
Every "give our AI agents memory" initiative treats memory as one thing to build. It isn't. This section works from four distinct memory types, borrowed from cognitive science and organizational-memory research, and follows what each one actually requires — governance, tooling, and honest limits included.
The Section as a Map
The same material as the cards below, arranged by what builds on what rather than by category. The four memory types carry their own colour; everything else is spine — narrative or cross-cutting. It opens on the core structure; the eight case studies come in when you ask for them.
Every page in this section:
- If Only We Had a Brain
- What This Adds Up To
- The Company Brain Is Four Different Problems
- Working Memory: The Stuff That Should Disappear
- Episodic Memory: Precedent, Not Law
- Semantic Memory: Expensive to Get Wrong
- Procedural Memory: How, Not What
- A Company That Doesn't Exist, Mapped Anyway
- What Done Looks Like
- When the Memory Lies
- When Two Agents Remember Differently
- Naming a Steward Isn't Enough
- When the Agent Is the One Writing the Memory
- Baked In, or Looked Up
- The Answer That Sounds Too Sure
- Operation Scarecrow
The Front Door
If Only We Had a Brain
An organization chasing a brain it doesn't have yet — the same way the Scarecrow wanted one, only to find out it had been doing the reasoning the whole time. Why Operation Scarecrow exists, and what it holds itself to while it's being built in public.
The Company Brain Is Four Different Problems
Every "give our AI agents memory" initiative treats memory as one thing to build. Cognitive science, three decades of organizational-memory research, and current AI-agent theory all say otherwise.
What This Adds Up To
Fourteen pages of argument, compressed into what to actually do, roughly in order, and what's still genuinely unresolved. For the reader who isn't going to read the other fourteen.
The Four Regions
The four memory types aren't a borrowed metaphor — each one is a claim about a different part of the brain, and they behave differently because the anatomy does. Working memory sits on the surface, in the prefrontal cortex. Episodic memory is buried in the hippocampus. Turn the model, or pick a structure by name, and each region will show you what this section has on it.
The four memory types and the structures they map onto:
- Working Memory — prefrontal cortex
- Episodic Memory — hippocampus
- Semantic Memory — anterior temporal lobes
- Procedural Memory — basal ganglia and cerebellum
Working, Episodic, Semantic, Procedural
The Stuff That Should Disappear
Decision queues, in-flight status, what a fleet coordinator is waiting on right now. Cheap, disposable, and the wrong place to put any governance at all.
Precedent, Not Law
Every dated, reasoned decision your org makes belongs somewhere — genuinely useful to the next person who hits the same question, but not yet a fact anyone should treat as settled.
Expensive to Get Wrong
The glossary, the entity model, the hard constraints — settled facts with the widest blast radius in the system. Why this tier needs a real steward, not just a good intention.
How, Not What
Skills, runbooks, and code — the one tier your engineers already govern well, and the one place AI writing its own memory still underperforms not having a memory at all.
Seeing It Assembled
A Company That Doesn't Exist, Mapped Anyway
Walking a fictional software org's actual artifacts — budgets, incident reports, runbooks, glossaries — through all four memory types, including the fifth case that refuses to sort into just one.
What Done Looks Like
Not a product to buy — an org design and an assembled architecture, most of which already exists in parts, with one piece that still has to be built from scratch.
Where It Gets Hard
When the Memory Lies
Three named attacks already target AI agent memory specifically, and the process meant to catch them can itself be gamed. What blast radius looks like per tier, and the one defense actually shown to work.
When Two Agents Remember Differently
A multi-agent fleet doesn't have one memory, it has many — and they disagree. Four ways that goes wrong, and why the fix depends on which tier the disagreement lands in.
Naming a Steward Isn't Enough
Roughly 80% of data-governance initiatives fail for the same reason: a role gets created, nobody gets actual authority. What a steward role needs to survive contact with a real org.
When the Agent Is the One Writing the Memory
Letting AI propose facts and skills instead of just reading them changes the governance problem completely — including one finding that undercuts the tidiest version of this idea.
Baked In, or Looked Up
Fine-tune a fact into a model's weights, or keep it in an external store it can retrieve. One reasons better. The other is the only one this whole framework's governance model can work with.
The Answer That Sounds Too Sure
A cleaner AI interface can make people check its answers less, not more — and a real court ruling has already held a company liable for a confidently wrong AI summary.