Skip to content
Memory SodaMemory Soda

Playground

An interactive console that drives the real /v1 API and shows every request and response as it happens.

It is a network inspector purpose-built for the memory pipeline. If you want to understand why something was or was not extracted, start here.

Setup

Section titled “Setup”

The playground authenticates with an API key, not your dashboard session, because it exercises the same surface your application does.

  1. Create a key on API Keys.
  2. Paste it into the key field at the top.
  3. Accept the generated dataset (ds_a1b2c3d4) or type your own.

Using a throwaway dataset keeps experiments out of real users’ memory.

Layout

Section titled “Layout”
┌────────────────────────────┬──────────────────────────────┐
│                            │  Ops │ Episodes │ Recall │ Facts │
│      chat transcript       ├──────────────────────────────┤
│                            │                              │
│                            │   selected tab               │
│                            │                              │
├────────────────────────────┤                              │
│  message input             │                              │
└────────────────────────────┴──────────────────────────────┘
  working / episodic / semantic settings panels

The ops log

Section titled “The ops log”

The centrepiece. Every API call the playground makes is recorded with:

  • operation type, thread_created, message_added, prepare, recall, auto_compacted, episode_scheduled, facts_extracted, error, …
  • relative timestamp, +0.0s, +1.4s, +12.8s from the start of the thread
  • duration
  • the full request body, response body and HTTP status

A typical turn:

+0.0s   thread_created      POST /v1/threads                          201   18ms
+0.4s   message_added       POST …/threads/f2cb…/messages             201   11ms
+0.4s   prepare             POST …/threads/f2cb…/prepare              200   24ms
+0.5s   recall              POST /v1/memory/recall                    200  318ms
+2.9s   ai_replied          POST …/threads/f2cb…/chat                 201 2410ms
+13.1s  episode_scheduled   (background)
+31.7s  facts_extracted     GET  …/semantic/…/facts?episodeId=8b21…   200   14ms

That last line is the one to watch: it appears when the pipeline finishes and tells you exactly which facts came out of that episode.

The log is in-memory and ephemeral. Reloading the page clears it, and there is no export.

Watching extraction

Section titled “Watching extraction”

useExtractionPoller polls for new episodes every 2.5 seconds (giving up after 90) and, when one completes, fetches the facts it produced and emits a facts_extracted op.

So the normal loop is:

  1. Send a few messages.
  2. Stop typing.
  3. End the thread, or wait out autoEpisodeIntervalMs (default 30 min) plus the scheduler tick.
  4. Watch facts_extracted land, ~20–60 seconds later.

The polling GETs are deliberately not logged as ops, they would drown the signal.

To skip the wait, use the End thread action, which calls POST /v1/threads/:id/end and queues extraction immediately.

Tabs

Section titled “Tabs”

Ops

Section titled “Ops”

The log described above. Click any entry to expand the full request and response.

Recall

Section titled “Recall”

Run recall() by hand against the current dataset, with all the controls exposed:

Control Maps to
Query query
Min confidence minConfidence
Limit limit
Include episodes / synthesis / raw include
As of asOf

The rendered context block is shown exactly as your application would receive it. This is the fastest way to see what your prompt is actually going to contain.

Facts

Section titled “Facts”

Live facts for the dataset, refreshed as extraction lands. Individual facts can be deleted, which emits a fact_deleted op.

Episodes

Section titled “Episodes”

Episodes for the dataset, with summaries, key learnings and status. Failed ones can be retried.

Settings panels

Section titled “Settings panels”

Working, episodic and semantic settings for the session.

Panel Notes
Working memory autoCompactThreshold, messageLimit, applied to threads created here
Episodic autoEpisodeIntervalMs, contextEpisodes, weights
Semantic factsInContext, thresholds

These are playground-local. They do not change project settings, and they reset on reload. To change real defaults use Project Settings.

Lowering autoEpisodeIntervalMs to a second or two makes experimenting much faster, thread-level overrides accept values down to 1000. Or end the thread; that fires immediately.

The chat panel

Section titled “The chat panel”

Messages are sent through POST …/chat, which runs the whole turn server-side against Gemini.

This endpoint exists for the playground. Your application should use prepare + recall with your own model. The SDK deliberately does not expose chat.

Each message shows its sequence number, timestamp, and any token/model/latency metadata. There is a manual message form for injecting system and tool roles and arbitrary metadata, useful for testing how extraction handles non-user turns.

Replies are not streamed; they appear when the call completes.

What it is good for

Section titled “What it is good for”
Question How
Why wasn’t that extracted? Send it, wait for facts_extracted, inspect the response
What will my prompt contain? Recall tab, read the rendered context
Is the pipeline working at all? Ops log shows every call and its status
How long does extraction really take? Relative timestamps between message_added and facts_extracted
Does raising factsInContext help? Change it in the semantic panel, re-run recall, compare
What does a contradiction do? State something, then contradict it, watch the old fact become superseded

Limitations

Section titled “Limitations”
Ops log is ephemeral No persistence, no export, no sharing
Needs an API key pasted in Cannot use your dashboard session
Settings are local Do not affect the project
No streaming Replies arrive all at once
Facts/Episodes tabs duplicate Datasets Same data, narrower view
Gemini only The chat endpoint hard-codes it

Next

Section titled “Next”