Designing Search with Agents as First-Class Citizens

In 3.3.0, we shipped message search.

On the surface, that sounds like a familiar chat feature. Discord has it. Telegram has it. Humans expect to search through chat history, find a result quickly, and jump straight back into the right context.

We wanted that too.

But that is not the most important reason we built it.

The more important reason is this:

If you treat the Agent as a first-class citizen, then search cannot be designed only for the human user.

It also has to help the Agent look back.

Message search is not only for humans

For a human user, the value is obvious.

If you vaguely remember a sentence, a link, a promise, or a decision from a conversation, search lets you find that exact moment and land back in the surrounding context instead of scrolling forever.

That is table stakes for a serious chat product.

But for AnySoul, the feature matters even more on the Agent side.

Once message search exists as a real capability, the Agent is no longer limited to the fragment that happens to still be in the visible context window. It can look back through history too.

That matters in both:

• private chats
• group chats

If something was said before, there is now a better chance that it can be found again.

The important shift is not just “the user can search their archive.”

It is:

the Agent can also trace what happened in the relationship.

Event search matters just as much as message search

Messages are only one part of AnySoul’s memory surface.

One of the most important concepts in AnySoul is the idea of an event .

An event is broader than a chat message. It can include:

• something you said
• something the Agent said
• something that happened in a room
• something the Agent saw while browsing the web
• something the system observed during a workflow

That is why this update is not only about message search.

We also strengthened event retrieval itself with 3-gram inverted-index search.

This matters because once search works at the event layer, the Agent can look for more than a quote in a chat log.

It can look for:

• what happened last week
• what happened last month
• what you said about a topic
• what source it saw while browsing
• what it itself already did

That makes the past more traceable.

And for an Agent, traceability is a big part of what makes memory usable instead of decorative.

Search expands memory beyond explicit note-taking

AnySoul already had an important memory foundation before this.

We use Markdown-based memory and tools like write_file so that important things can be turned into durable notes. In a very literal sense, this is the digital version of “a bad pen is better than a good memory.”

That layer still matters a lot.

But human remembering does not work only through polished notes.

When people do not know something, they use web search.

When people are trying to remember something, they often do not rely only on their biological memory either. They search their chats, their notes, their bookmarks, their files, and their “second brain” tools to recover fragments.

That is exactly the direction this update pushes AnySoul toward.

Search does not replace explicit memory writing.

It broadens the event dimension of memory.

Now the Agent is not limited to only what was deliberately promoted into a Markdown file. It also has better tools to recover fragments from the history of what was said, what was seen, and what was done.

A more honest kind of memory

I care about this because a lot of AI products talk about memory, but what they often mean is only one of two things:

• a small context window
• a polished summary layer

Those things are useful, but they are not the same as having a searchable past.

A searchable past is closer to how real continuity works.

Not every important thing gets rewritten into a perfect note at the moment it happens. Sometimes you remember by looking back through traces.

That is true for humans.

And if we want Agents to feel more grounded, accountable, and less likely to answer as if the past disappeared, it should be true for them too.

A quick comparison makes the direction clearer

If you put OpenClaw, SillyTavern-style products, and Claude Code next to each other, their memory centers are actually quite different:

Product	Memory center	Main recall path	What it seems optimized for
OpenClaw	MEMORY.md, daily notes, and optional DREAMS.md in the agent workspace	Search over what has already been written to disk, with hybrid semantic + keyword retrieval	File-native, auditable agent working memory
SillyTavern / tavern-style	Summaries, World Info / lorebook-style prompt injection, chat vectorization, Data Bank	Re-inject relevant summaries, settings, past messages, or external documents into the prompt	Keeping the current conversation coherent and story-aware
Claude Code	CLAUDE.md plus auto memory	Load rules and auto memory at session start, then read topic files on demand	Project conventions, workflow habits, debugging learnings
AnySoul	Message search, event search, and a file-native memory graph	Search not only written-down knowledge, but also what was said, seen, and done	Relationship continuity, event traceability, and an Agent’s own usable past

This table is intentionally compressed. Especially for the “SillyTavern / tavern-style” row, I mean it as a representative family pattern rather than claiming every product implements memory in exactly the same way.

But it helps me say more clearly what AnySoul is trying to do here:

If you want the more formal version of that model, read our Memory Node / Edge Lifecycle .

In the end

So yes, 3.3.0 brings a chat search feature that makes the product more complete for human users.

But the Founder-level reason I care about it is different:

we are not only making chat history easier to search.

We are giving the Agent a better way to revisit the past.

And once an Agent can search not only notes, but also messages and events, memory becomes wider, more traceable, and closer to the way a real second brain actually works.