What Your AI Character Actually Remembers - A Plain Guide to Reverie's Memory System

"It forgot." It's the most common frustration with AI chat, and almost always the most misunderstood.

Your AI character did not forget the way a person forgets. It either (a) never had the information in the context window when generating the reply, or (b) had it but didn't reach for it. These are different problems with different fixes.

This is a plain-English explanation of how memory actually works in Reverie, what survives a long arc, what doesn't, and what you - the user or creator - can do about it.

The three layers of memory

Reverie's character memory is not a single bucket. It's layered, and each layer behaves differently:

Conversation context - the recent messages the model can directly see when it generates a reply. Bounded by a token budget. Oldest messages roll off.
Conversation summary - when a thread gets long, the system writes a compressed narrative summary of the older parts and feeds that into context instead of the full transcript. You don't lose the storyline, but you do lose verbatim wording.
Long-term memory - durable facts about you and the relationship that persist across sessions: name, key personal details, preferences, important events, relationship milestones.

When your character "remembers your birthday" three weeks later, that's long-term memory. When your character can still talk about a fight you had earlier this session but not the exact line you said, that's the summary doing its job.

What gets remembered automatically

Reverie generates long-term memory entries on its own. You don't have to do anything for the common stuff to stick:

Personal details - your name, age (if mentioned), where you said you live, what you do.
Preferences - things you said you like or dislike, in the world of the conversation or about the character.
Relationship milestones - first kiss, falling out, reconciliation, declarations.
Character growth - shifts in the character's own state you've established (they got the job, they quit drinking, they finally called their mother).

You can browse, edit, and delete these manually. This is the single highest-leverage feature most users never open. If your character "keeps getting your name wrong" or "forgot we broke up," you can fix it in one minute by editing the memory directly instead of trying to re-explain it in dialogue.

What automatic summarization preserves (and what it sacrifices)

When your conversation grows past Reverie's message or token thresholds, the system writes a narrative summary of the older parts. The summary is designed to preserve:

The current state of the relationship.
Emotional dynamics ("they're still angry about X").
Open-ended plans and unresolved threads.

And designed to sacrifice:

Verbatim wording. The character won't quote your line from 200 messages ago.
Exact dates, specific numbers, fine logistical detail.
Most secondary characters mentioned in passing.

The practical implication: if a specific line, promise, or detail matters for what comes later, don't trust the summary with it. Pin it as a long-term memory entry yourself. The 30 seconds you spend doing that buys you 50 messages of continuity.

Global vs. character-specific memory

Reverie has two scopes:

Character-specific memory - what this character knows about you. The default.
Global memory - facts that apply across every character you interact with (e.g., "uses they/them," "is allergic to seafood," "is writing a novel").

Global memory is for facts that shouldn't depend on which character you opened. Use it sparingly and for things that are about you, not about a specific relationship. Don't put "we kissed in chapter 4" in global memory - that's a per-character, per-arc thing.

Identities are their own memory bucket

Reverie's user identities feature is more powerful than it sounds for memory specifically: each identity has its own memory with a given character.

If you talk to one character as "Mira the bartender" in one arc and "Captain Eun" in a sci-fi RP, the character remembers the two of you as two different people. No bleed.

This is the cleanest way to run multiple roleplays with the same character without one storyline poisoning another. It also lets you experiment safely - blow up a relationship in one identity, the other is untouched.

Forking has its own memory

If you fork a conversation, the new branch carries the memory that existed at the fork point - and from there, builds its own. The two branches don't merge memory back together. If you fork to try an alternate scene, the character in the original branch never "learns" what happened in the alternate.

This is a feature, not a bug. It's what makes forking useful for trying scenes without consequences.

How to make sure something gets remembered

If a moment really matters, four things stack the odds:

Reference it again within the next few turns. Recency keeps it in raw context longer, giving the summarizer a chance to weight it.
Put it in your own words back to the character. "So you really did mean it when you said you'd come back." Now it appears in your message, which is hard to overlook.
Pin it as a manual memory entry. Open memory management and add a one-line note. This is the durable fix.
For creators: add it to the character's scenario or persistent context. (Scenarios & memory guide.) Anything in the scenario survives forever; it's part of the character, not a memory.

How to make sure something doesn't get remembered

The flip side, equally important. Sometimes the character keeps bringing up a thing you wish was buried.

Edit the offending memory entry. Direct fix.
Don't re-mention it. Even denial counts as a reference; the summary will weight it.
For one-off experiments, fork. The branch is sandboxed; whatever happens there doesn't follow you back.

Common memory mistakes

Repeating yourself instead of pinning. If you've had to remind your character of the same fact three times, the issue isn't reminding harder - it's that the fact never made it into long-term memory. Open the panel.
Assuming the AI knows what you know. The model can't read minds, can't see what you didn't say. If a backstory beat happened off-screen, write it as a memory entry.
Confusing "forgot" with "drift." Long arcs sometimes have characters slowly shift voice or stance. That's usually not memory loss - it's the model averaging across a longer summary. Lower the temperature, or restate the character's stance once in dialogue.
Putting plot in global memory. Plot is per-character. Global is for you.

How this stacks with the rest of Reverie

Conversation forking - sandboxed memory for experimental scenes.
Scenarios - things that should never be forgotten go here, not in chat.
Identities - run parallel arcs cleanly.
Good character writing - a character with a distinct voice and a clear "Won't" list (guide) drifts less because there's less ambiguity to average across.

The mental model

Stop thinking of AI memory as "does it remember." Start thinking of it as: what does this character have in front of them right now, and what facts about us do they carry between sessions?

The first one is bounded and rolls. The second one is editable, and most users never touch it.

The hour you spend understanding which is which is the hour your long arcs stop falling apart.