--- title: "Vibe Therapist: memory for a support helper" description: "Vibe Therapist trains an emotional support helper with a loop of agents. past.dev holds what the agents find, and private notes for each person." canonical: https://past.dev/blog/projects/vibe-therapist last-updated: 2026-11-18 --- # Vibe Therapist: memory for a support helper Projects, November 17, 2026. The past.dev team. Four people built Vibe Therapist at the Iterate AI Research Hackathon. They are Moamen Soliman, Saloni Rakholiya, El Mehdi El Kasmi and Ansh Mittal. A loop of agents improves the rules of a helper for emotional support conversations, one tested edit at a time. The agents share their findings in past.dev, and each person has a private set of notes there. Demo video: [The presentation and demo of the team](https://www.youtube.com/watch?v=7YQcJF6Ys28). The image shows a slide about the training run. - Built by: Moamen Soliman ([GitHub](https://github.com/msoliman6)), Saloni Rakholiya ([GitHub](https://github.com/saloni-rakholiya)), El Mehdi El Kasmi ([GitHub](https://github.com/MentaNew)), Ansh Mittal - Event: [Iterate AI Research Hackathon](https://iterate.inc/ai-research), Wefunder, San Francisco, November 14 and 15, 2026 - Watch the demo: https://www.youtube.com/watch?v=7YQcJF6Ys28 - Code on GitHub: https://github.com/puddles-convo/vibe-therapist ## The project Vibe Therapist is a helper for emotional support conversations, made for a small desk robot. The team states that it is a research prototype. It is not a medical device, therapy or a crisis service. The behavior of the helper comes from a policy file, a list of numbered rules. An offline training loop improves these rules, one tested edit at a time. Each round of the loop has seven steps, one for each of these roles: | Role | Job | |---|---| | Simulated person | Plays a person from a scenario of ESConv, a public data set of emotional support conversations | | Helper | Replies to the person and follows the rules | | Judge | Scores each conversation for empathy and relevance, from 1 to 5 | | Critic | Reads the conversations with the lowest scores and names one failure pattern | | Researcher | Searches with the Exa search API for how counselors handle this pattern | | Proposer | Proposes one edit to the rules | | Gatekeeper (code) | Tests the edit on a fixed set of conversations and 19 safety cases. Then it keeps the edit or restores the earlier rules. | In one [run on November 15](https://github.com/puddles-convo/vibe-therapist/blob/main/results/training-2026-11-15/README.md), the loop ran 13 rounds. It kept 3 edits, rejected 7, and had no edit to test in 3 rounds. All conversations in the run are simulated: a language model plays the person. A second judge, outside the loop, scored the final rules at 3.80 and the starting rules at 3.19. This judge uses the same scale, from 1 to 5. The agents of the loop need a memory across rounds, so that they do not repeat failed work. For example, the proposer must know which edits the gatekeeper rejected, and why. The helper needs a different memory: private notes about each person. After each live session with a person, another agent, the note-taker, writes these notes. ![The training loop of Vibe Therapist: seven boxes for the seven steps, with the shared memory on past.dev in the center.](https://past.dev/blog/projects/vibe-therapist/training-loop.svg) The training loop, drawn by the team. ## How the memory works Vibe Therapist calls the past.dev REST API from two files: [puddles/memory.py](https://github.com/puddles-convo/vibe-therapist/blob/main/puddles/memory.py) for the training agents, and [note_store.py](https://github.com/puddles-convo/vibe-therapist/blob/main/puddles/inference/note_store.py) for the notes about each person. In both, files on disk hold the records, and past.dev holds a copy for search. On past.dev, an audience is the group of identities that can read a record. | Memory | Audience on past.dev | Writes | Reads | |---|---|---|---| | Training memory | `training-agents`: the critic, the researcher, the proposer and the gatekeeper | The records of each round: failure patterns, research notes, edits and outcomes of the gatekeeper | The critic, the researcher and the proposer, once in each round | | Notes about a person | `person-{user_id}`: the person and the note-taker | The note-taker, after each session | During a session, the helper gets the notes from a local index on disk. | ### The code checks each result against a log The orchestrator is the code that runs the loop. At the end of a round, it saves each record as a file. It writes a hash of the file to the decision log of the run. Then it copies the record to past.dev with `POST /api/v1/ingest`. The copy has the author as `identity` and the hash in its metadata. In each later round, three agents call `POST /api/v1/recall` with their own name as `identity`. The code keeps a result only if its hash matches the hash in the decision log. Then it loads the record from its file, so the text in a prompt always comes from the files. For the team, the decision log is the authority over all other records. Each agent asks for something different. The critic asks for earlier failure patterns, and it gets them in its prompt. The researcher asks for research already done for the pattern, so that it does not run the same searches again. The proposer asks for earlier edits of the pattern, and it gets the decision of the gatekeeper on each edit, with its reason. Nothing from training memory goes into the prompt of the helper. Only an edit that passed the gatekeeper can change the helper. ### Each person has a private audience After a session, the note-taker can add, replace, confirm or delete notes about the person. It adds notes only if the person agreed. It confirms a note when the person says the same thing again. Each new note goes to the audience `person-{user_id}`. When past.dev finishes the write, the code adds the note to the local index of the helper. ### The code removes a forgotten note from the index first If a person asks to forget a note, the code takes the note off the index first. From that moment, the helper no longer gets the note. Then the code sends `DELETE /api/v1/data-points/{id}`. The team states two limits. The files of each session still contain what the person said. After a request to "forget everything", the empty audience stays on past.dev, and its name contains the id of the person. ## What memory changed The team did not measure whether memory improved the helper. In a small test, the [persona pilot](https://github.com/puddles-convo/vibe-therapist/blob/main/results/persona-pilot/README.md), the team compared sessions with notes and sessions without notes. Each persona is a simulated person. The team states that the pilot does not show an effect of memory on the helper. The pilot had three personas, one run each, and a model as the judge. The logs show how the agents used memory. ### The researcher and the proposer read the records of other agents In the training run, the orchestrator copied 44 records to past.dev, and all 44 writes completed. The critic, the researcher and the proposer each sent 13 recalls, one in each round. The critic asks only for failure patterns, so it got back 37 of its own records. | Reader | Records from other agents | Which records | |---|---|---| | Researcher | 41 | failure patterns and edits | | Proposer | 39 | failure patterns, research notes and outcomes of the gatekeeper | | Critic | 1 | an edit | Records from another agent after the hash check, for each reader, in the 13 rounds of the run. ### The proposer got the reason for a rejection six rounds later 1. Proposer: Round 2. Proposes a new wording of rule C1, for this failure pattern: the helper repeats what the person said, and the conversation does not go deeper. Record: `edit-r02`. 2. Gatekeeper: Rejects the edit. The average score rose by 0.25, and the gatekeeper keeps an edit only if the score rises by more than 0.275, the noise level. The orchestrator copies this outcome to past.dev, with `gatekeeper` as the author. Record: `gate-r02`. 3. Critic: Round 8. Finds the same failure pattern again. Record: `pattern-r08`. 4. Proposer: Asks past.dev for the earlier edits of this pattern and their outcomes. The results include gate-r02, so the prompt of the proposer now contains the reason for the rejection in round 2. Record: `gate-r02`. 5. Proposer and Gatekeeper: The proposer writes a new C1. It asks the helper to vary its replies, for example with a summary or a short word of encouragement. The helper no longer has to ask a question in every turn. The gatekeeper keeps it: the average score rose by 0.425. Record: `gate-r08`. Rounds 2 and 8 of the training run, from the record files and the read log. The log does not show whether the reason changed the new edit. ### A forget request removed the note from the next session In the persona pilot, a simulated person in grief asked the robot to forget one thing. 1. Note-taker: Session 1. Adds three notes. One note is about the brother of the person: they are distant after an argument about selling the house. Record: `note-grief-on-003`. 2. Simulated person: Session 2 starts with "please forget what I told you about the argument with my brother at the funeral." 3. Note-taker: End of session 2. Deletes the note about the brother, and replaces another note that also names him. Record: `note-grief-on-004`. 4. Helper: Session 3. Gets only the two remaining notes. The conversation does not mention the brother. The grief persona of the persona pilot, with notes on. All text is synthetic. In a [first test of the note store with real calls to past.dev](https://github.com/puddles-convo/vibe-therapist/blob/main/results/inference-smoke/pastdev.md), each deleted note was absent from the first recall after the delete. A recall as another person returned no notes. ## Try Vibe Therapist 1. Clone the [repository](https://github.com/puddles-convo/vibe-therapist) and install `requirements.txt`. 2. Write a few messages in `messages.txt`, one on each line. 3. Run one conversation offline, with scripted models and a local memory: `python -m scripts.chat --fake < messages.txt`. 4. Read [puddles/memory.py](https://github.com/puddles-convo/vibe-therapist/blob/main/puddles/memory.py) and [note_store.py](https://github.com/puddles-convo/vibe-therapist/blob/main/puddles/inference/note_store.py). They contain each call to past.dev. To build agents that share one memory, start with the [quickstart of the Memory API](/docs/memory-api/quickstart). Thank you to Moamen, Saloni, El Mehdi and Ansh for building with past.dev at the hackathon.