Comparison

Coldcache vs Perplexity

Same category — live web search, cited sources, streaming answers. The difference is what happens to a Space when you edit it, and whether a thread is an artifact or a scrollback.

Three differences that matter

Pinned versions vs. in-place mutation

PERPLEXITY

A Space is a mutable singleton. Edit its instructions and every thread that already ran now resolves against the new prompt. There is no record of what the old one said.

COLDCACHE

Editing publishes v2. v1 is frozen forever, still runnable, still linkable. A chat stores agent_version_id, so publishing v4 on Monday does nothing to the chat you opened Friday on v3.

You can reproduce any answer this product has ever given you, including the configuration that produced it.

Declared state vs. accumulated context

PERPLEXITY

Spaces accumulate files and context, and threads carry their own history. Useful for exploration; fatal for a controlled comparison.

COLDCACHE

Every run loads the pinned version into a fresh context window. Memory is declared inside the version body, so turning it on is a version bump you can see in the diff and in the chat header.

State is a feature you opt into and can point at, not an ambient condition.

Run records vs. threads

PERPLEXITY

You get a thread. To attach it to a ticket you screenshot it.

COLDCACHE

You get a run record: input hash, version id, resolved model id, every tool call, every source with a fetch timestamp and content hash, output, output hash, latency, tokens, cost. Diffable, exportable, addressable.

The unit of work is an auditable record, not a conversation you have to screenshot.

When to use which

Use Perplexity

You are exploring. One person, one question, read it and move on. Perplexity is genuinely good at this and often faster, and it has spent far more on the index than we ever will. If nothing downstream depends on the answer being the same tomorrow, that is the right tool and we will not pretend otherwise.

Use Coldcache

The same question gets asked every week, by a team or a scheduled job. Someone will eventually ask why the answer changed, and “the model must have updated” is not an acceptable reply. You need the exact configuration that produced it, frozen, diffable, and controlled — so that not everyone can quietly change the thing your process depends on.

Side by side

	Perplexity	Coldcache
Live web search + citations	Yes	Yes
Edit instructions	Mutates in place	Publishes a new version
Old configuration retrievable	No	Forever
Chat bound to a specific version	No	Permanently, for the life of the chat
Memory across sessions	Space context accumulates	Off by default, declared per version
Same input across two configs	Manual, one at a time	Side-by-side diff, batch across an input set
Web held constant for a comparison	Not possible	Replay mode re-serves stored snapshots
Choose your model provider	Provider menu, their account	Claude, ChatGPT or Gemini, named per version
Export a run as evidence	Copy the text	JSON with hashes, sources and settings
Who can change an agent	Anyone the Space is shared with	Per person, per agent: read, write or approve

A Space is a place to think. A pinned version is a thing you can be held to. Once an answer leaves your building — to a client, an auditor, a regulator — you need the second one, and no amount of index quality substitutes for it.

Compare with ChatGPT

Pin an agent. Run it in a year. Get the same thing back.

Claude, ChatGPT and Gemini included — no provider account to open. Build one agent, publish v1, run it, then change the prompt and diff the two.

Create a workspace Read the quickstart