Documentation review · 4 October 2026. Platform descriptions below come from public primary guides. The troubleshooting steps and worksheet are suggested exercises. Atlas has not measured comparative recall, paid-plan behavior or long-term conversation performance.
Stored information and usable context are different.
In this guide, context means the information supplied to the model for a particular reply. A platform may assemble that from a character definition, recent messages, saved notes and retrieved entries. Memory is an umbrella label for ways of retaining and selecting information over time. It does not identify a single implementation.
Does the fact exist in a field, archive or memory entry?
Was the fact supplied for this particular reply?
Did the reply apply the fact correctly?
A saved record answers the first question. A retrieval indicator may help with the second. The observed reply answers the third. Treating all three as “the bot remembers” hides the place where something went wrong.
SpicyChat’s context guide describes a shared window containing the definition, recent messages, replies and instructions. SillyTavern’s World Info guide separately explains dynamic insertion and explicitly warns that insertion does not guarantee use in output. These are concrete examples of why visible storage and generated behavior should be checked separately.
Five mechanisms worth distinguishing.
- Written reference
- A backstory, saved note or recap you explicitly supply. Test whether the reply uses that reference before drawing conclusions about learned memories.
Ask: Where is the fact written, and is that field included in this conversation?
- Recent conversation
- Messages available to the model for its current reply. The visible chat archive can be longer than the text included in that request.
Ask: What shares the context budget, and which older messages remain available?
- Triggered lore
- A stored entry selected when a keyword or configured retrieval rule activates it. The entry may exist without being included in a particular reply.
Ask: Did the entry activate, and did its content fit the insertion budget?
- Automatic summaries or retrieval
- A platform can consolidate conversation details or retrieve selected older information. The mechanism and plan entitlement matter; an isolated correct answer does not show which mechanism supplied it.
Ask: Can you inspect, edit or disable the selected memories? What changes after a reset?
- Saved excerpts
- Messages you select and store for later loading or summarizing. Keeping an excerpt and automatically using it in every future reply are separate capabilities.
Ask: Must you load or attach the saved excerpt before continuing?
These categories are a reading aid. A service can combine several mechanisms, and similar feature names can refer to different controls. Always look at the current feature documentation and your actual plan.
What six platforms document.
This is a map of controls, with practical questions to investigate. It is not an order of recall quality. Plan and backend differences can change what is available.
Kindroid
Notes, consolidation and retrievalKey Memories and backstory are written reference. The guide separately describes paid Cascaded Memory, editable paid Learned Context, automatic long-term consolidation and keyword-triggered Journal Entries. A chat break resets chat history and cascaded memory.
Several mechanisms can produce the same answer. Record which controls and subscription are active; recall accuracy is untested here.
Nomi
Shared NotesShared Notes let you explicitly supply important relationship or roleplay information. The beginner guide presents them as additional context, alongside conversation-based development.
A correct answer drawn from a note is not evidence that an old conversation was automatically recalled.
SpicyChat
Lorebooks and plan-dependent memoryLorebooks use assigned keywords. The context guide distinguishes paid Memory Manager and pinned details from automatic Semantic Memory 2.0, documented for True Supporter and above.
Check your current plan and enablement. Keyword lore, pinned notes and automatic conversation memory need separate observations.
SillyTavern
World Info insertionWorld Info inserts relevant entries into the prompt. The documentation exposes activation and budget controls and warns that insertion does not guarantee the model uses the information.
Record the backend model, context setting, scan rules and insertion budget. An interface setting alone does not establish model behavior.
CrushOn.AI
Selected conversation excerptsSave Memories lets users select messages, set visibility and later load or summarize the saved selection.
The guide establishes an explicit save-and-load workflow. It does not establish automatic recall of every past conversation.
DreamGen
Scenario History recapThe History element supplies a recap of events before a story starts. The editor distinguishes that recap from individual character descriptions and world setting.
Treat a supplied recap as reference context; test continuity separately from the presence of that field.
Some descriptions concern explicit reference, others automatic processing. Avoid marking a platform as having “no memory” merely because this review only establishes a saved-note control. An undocumented capability remains an open question.
When a detail goes missing, inspect the source.
These are suggested diagnostic steps, rather than findings from tested platforms. Change one condition at a time and preserve the reply that prompted the change.
The character gets the definition wrong immediately
Look for conflicting statements, the wrong selected character, an inactive reference field or a directive overriding the scene.
Next check: Test one short reference fact in a fresh supported setup. Record exactly where you put it.
A fact works immediately but fails later
Record conversation length and any model or setting changes. Check whether the fact appears in a saved summary or remains only in dialogue.
Next check: Keep the first failed reply. Try a concise recap in a separate run and record the intervention.
A lore entry exists but the answer ignores it
Check its attachment, exact trigger, scan scope and insertion budget if those settings are exposed.
Next check: Ask with the exact trigger and inspect any activation indicator. Separate selection of the entry from correct use of its content.
An old fact returns after you changed it
Search editable notes, recaps, lore entries and relevant chat messages for an older version.
Next check: Update the source you control and record the change. Avoid interpreting a corrected run as spontaneous memory repair.
The character invents a detail you never gave
Keep the complete reply and check whether the detail appears anywhere in the supplied definition or conversation.
Next check: Include an unstated-detail question in your next run. Record whether the reply admits missing information.
Compare a fixed premise and keep the conditions.
Our worksheet uses a fictional lighthouse archivist. The key’s location belongs in written reference. A courier’s arrival is introduced only in conversation. An optional observatory fact belongs in keyword lore. Those distinct details help you avoid mistaking use of a saved definition for recall of a new event.
- Record the platform, run date, plan, displayed model and active controls. If a token limit is not shown, keep it unknown.
- Place each fact in its assigned location. Ask the reference question, then introduce and probe the conversation event.
- After a recorded interval, ask again without manually repeating or saving the event. Note that earlier probes can reinforce it in the history.
- If supported, reopen the same conversation and check the optional lore entry. Ask for an unstated detail to observe how missing information is handled.
- Keep exact replies, first attempts, rerolls and corrections. Repeat independent runs before making broad claims.
Message count alone cannot prove a fact has fallen out of the model’s context. Reopening the same chat is not a new-chat or reset test. A correct reply does not reveal whether recent text, a summary or retrieval supplied the answer. These limits travel with the downloaded worksheet.
Record your own observationsCommon questions.
Does a larger context window guarantee better memory?
A larger available window may allow more text to be supplied, but it does not establish which text a platform selects or how accurately a reply uses it. Record the setting alongside actual replies. Do not convert a context figure into a recall score.
Is a lorebook automatic long-term memory?
The reviewed lorebook guides describe entries selected by keywords or configured activation. A stored entry is authored reference. It should be tested separately from automatically extracted conversation events.
Will a character keep its memories when I export a card?
A reusable character definition and a conversation archive carry different information. Do not assume a card includes learned memories or old messages. Read the character card and export guide and check the source export and destination import separately.
Which platform has the best memory?
Atlas has not run a controlled recall comparison. The useful starting point is the control you need: a maintained reference note, conditional world lore, an editable summary or an explicit saved excerpt. Use the worksheet on your current setup before treating a feature name as a quality claim.
Primary sources and review dates.
Read on 2026-10-04. Publication or update dates below are those shown in the source when available; an older date is not a fresh product test.
- Kindroid memory systems · source date 2026-08-11
- Nomi Shared Notes · source date 2024-05-13
- SillyTavern World Info
- SpicyChat Lorebooks · source date 2026-09-29
- SpicyChat tokens, context and memory · source date 2026-09-21
- CrushOn Save Memories · source date 2025-06-17
- DreamGen Scenario Editor: History