Honcho vs MemoryRouter: Whose Memory Should This Be?

John Rood··4 min read

Two assistants talk to the same person. One hears that a project has been postponed. The other was absent. Should both know about the delay tomorrow?

For your personal coding assistants, carrying the decision over saves another explanation. For characters in a simulation, giving both the same knowledge can ruin the interaction. The user identity is identical in both cases. The memory requirement is not.

Honcho makes the observer part of its memory model. MemoryRouter's MCP connection gives a supported client access to the vault you selected. Choose between them by deciding whether a fact should travel with the person or stay relative to whoever witnessed it.

Honcho can remember a relationship from one side

A Honcho peer can represent a person or an agent. Its peer representation documentation describes two different observation settings.

With observe_me enabled, Honcho builds a representation of that peer from messages the peer has sent across sessions. This is enabled by default. It is Honcho's accumulated understanding of the peer.

observe_others enables something more specific at the session level. A peer forms representations of other peers from messages it observed in sessions they shared. An assistant that participated in Monday's conversation can have a different representation of you from one that only participated on Tuesday.

Those representations contain reasoning about the peer, including conclusions drawn from messages. They are more than copies of a transcript. But the distinction here is whose observations inform that reasoning, rather than how much text the system retains.

Consider a fictional user, Mara, and two agents:

Conversation Participants Information Mara provides
Monday Mara and the planning agent The launch has moved to December
Tuesday Mara and the writing agent The announcement needs a calmer tone

With peer observation configured for this setup, the planning agent's representation of Mara is grounded in their shared conversations. The writing agent has a different history to draw from. Asking what Honcho knows about Mara globally is not the same request as asking what the writing agent knows about her.

The current Peer Chat API exposes a target field for querying another peer's representation from the perspective of the peer addressed by the request. That is a concrete interface for the distinction, not a suggestion to put a character name in your prompt.

This is useful for simulations or assistants whose individual relationships are part of the product. Observation settings are not, by themselves, proof that your application's authentication and authorization are correct. Test those separately.

A shared vault deliberately removes that difference

I built MemoryRouter so a decision saved through one supported assistant remains available when you move to another. If both assistants are helping you ship the launch, the writing assistant should be able to retrieve the postponement even though you told the planning assistant first.

For MCP, authorization binds the connection to a selected vault. Authorize a second supported client against that vault, and its searches can reach the same saved context. It does not need to have attended the conversation that produced the memory.

The MemoryRouter MCP documentation defines this boundary explicitly. The search tool does not accept an observer identity or a parameter that selects another peer's viewpoint. Conversational project labels do not establish an access boundary. Data that must stay isolated belongs in distinct vaults, each with its own connection.

That gives you a different tradeoff from observer-relative memory. Separate vaults can isolate data, but they do not automatically construct two agents' different interpretations of the same person. A shared vault makes saved context available across its authorized clients, but relevance-ranked search does not load every saved fact into every answer.

The MCP setup page covers that shared connection. Generic MCP recall and saving remain model-directed. Connecting the second assistant does not import its old conversations or prove that either assistant has stored the launch decision.

Run the evaluation in both directions

Use the fictional conversations above rather than private customer data. For Honcho, configure the observation settings deliberately, then query Mara from each agent's perspective. Inspect whether the planning agent's view and the writing agent's view reflect their different shared sessions. Compare those with Honcho's own representation of Mara so you do not mistake the global view for an observer's view.

For MemoryRouter, save the December decision with store_memory from one connected client. Wait until a search finds it. Then authorize a second supported client for that vault and search for the launch date without including December in the query. Inspect the returned memory, not whether the assistant happens to guess the month.

Get a MemoryRouter vault for your assistants and verify that the launch decision travels between them.