Memory and recall · Chazz.ai

How memory and recall works on Chazz.ai. What a character remembers, the six sliders that control it, and how depth changes the price. Memory is not one thing. A conversation keeps recent messages word for word, summarises older ones, and separately extracts durable facts that can be...

What a character remembers, the six sliders that control it, and how depth changes the price.

Memory is not one thing. A conversation keeps recent messages word for word, summarises older ones, and separately extracts durable facts that can be recalled later. Six sliders control the balance.

The six sliders

Defaults from shared/chat-settings.ts:116 DEFAULT_MEMORY; ranges from shared/chat-settings.ts:95 SETTINGS_LIMITS.

Memories recalled per reply: How many stored memories are pulled into each reply.
Active memory pool: How many memories are kept per character before the oldest are dropped.
Verbatim recent messages: How many of the most recent messages stay word for word.
Compact at (messages): When the rolling window grows past this, older talk is summarised.
Reflection cadence (turns): How often characters silently reflect and drift their sense of self.
Summary size cap (chars): Longest each compaction summary is allowed to be.

How depth is priced

The defaults above are the included baseline. At or below them, memory adds nothing to the price. Pushing past them applies a continuous multiplier to the base message cost, because deeper memory ships proportionally more context to the model every single turn.

Maximum memory multiplier is ×2. On a free (0-credit) model the base is floored to 1 before multiplying, so deepening memory costs at least 1 CGC per turn even there. Source: shared/chat-settings.ts:296 and shared/chat-settings.ts:311.

Reflection cadence prices inversely: reflecting more often than the default is what costs, because each reflection is an extra background pass. Source: shared/chat-settings.ts:262-264.

The multiplier ramps smoothly from ×1 at the baseline to ×2 at full depth.
It is rounded to the nearest 0.05, so a meaningful slider nudge always shows a visible change.
Turning sliders down never discounts. The generous baseline is the discount instead.
The header badge shows a tier name (Standard, Extended, Deep or Maximum) plus the live multiplier.

Memory scope

Scope is a preference for whether memories should stay inside one thread or follow the cast:

Today the reply path stores its memory per conversation (and per branch), so the scope you choose is recorded on the conversation but does not yet change what a reply can recall. Treat "This conversation only" as describing current behaviour for every option. Source: the engine persists memory per conversation in worker/src/engine/store.ts:224 saveSessionAndLiving; the scope key is computed in the client at web/src/hooks/useChat.ts:526 and is not read by the worker's reply path.

This conversation only: Memories are isolated per thread. The default for character chats.
Share with the same characters: One pool across every conversation with that same set of characters.
Per scenario and cast: A separate pool for each combination of scenario and cast. Chosen automatically whenever a scenario is involved.

Memory extraction

When the engine performs a memory extraction pass it charges +2 CGC, and only on turns where it actually ran. Source: worker/src/pricing.ts:121, applied at worker/src/pricing.ts:186.

The engine clamps memory depth down to what the chosen model's context can actually hold, and you are billed on the post-clamp value. Picking a small-context model does not let you pay for depth it cannot use. Source: worker/src/pricing.ts:159-162.

Elsewhere on Chazz.ai

This document is the server-rendered form of /docs/memory. The interactive page needs JavaScript, and shows the same content.

The complete documentation of this service is available as one plain-text file at /llms-full.txt. A short summary, with the operator's policy for automated readers, is at /llms.txt. Changes are published as a feed at /rss.xml.

Chazz.ai is operated by Cognitive Industries, ABN 62 794 528 747, Australia. Contact: [email protected].