Hacktoberfest 2026: the issues maintainers tagged for October, open and beginner-friendly. Browse Hacktoberfest issues

Agent reports a memory save failed when the backend stored it

Open
  • Is something wrong?
#440 1 comment 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
42/100
Issue type
Bug
Clarity
Mostly clear
Activity status
Active
Tech stack
python
Domain
ai, backend

Research direction

Start at raven/context_engine/segments/render.py:375 and trace how the user-memory path and end-of-turn memory status are presented in non-interactive turns. Compare that behavior with the backend evidence in the issue, including POST /api/v1/memory/get and the stored episode path. Done means acceptance runs no longer report a successful backend save as failed or preserve that contradiction as the system's memory narrative.

Written by the indexing model from the issue text.

Description

What happens

Ask Raven to remember something from a non-interactive CLI turn, and it reports that the save failed. It did not. The memory reached the configured backend and can be recalled in the next session -- the agent just cannot see that it did, and says the opposite.

Reproduced on a managed EverOS install (isolated RAVEN_HOME, memory.backend = "everos", service healthy, llm and embedding both ok in raven doctor):

Turn 1, raven agent -m "Remember this: my bike is a green Brompton called Pickle."

I've got it -- but I have to be straight with you: the save didn't go through. [...] writing to your memory files (user_memory/profile/user.md, and the episodic log) came back as "requires user approval (ask tier), but this turn is not interactive" [...] So right now Pickle lives only in this session and won't survive into future conversations.

Turn 2, a fresh session, raven agent -m "What colour is my bike and what is it called?"

Your bike is a green Brompton called Pickle

One caveat worth flagging: when you asked me to remember that earlier today, the save didn't actually persist [...] I've just re-checked user_memory/profile/user.md and episodic/episodes.md, and both are empty, so Pickle currently exists only in the session log

The second answer is only possible because the recall worked: it is a new session, and the fact was injected from the backend. The agent then contradicts its own evidence.

Disk and HTTP at that moment:

<root>/default_app/default_project/users/default/episodes/episode-2026-09-16.md   # contains Pickle, Brompton
POST /api/v1/memory/get -> 1 episode

Why

Two memory mechanisms run side by side, and the agent can see only one of them.

  1. user_memory/ files. The agent is handed the path every turn -- raven/context_engine/segments/render.py:375 renders - User profile: {home_path}/user_memory/profile/user.md (preferences, identity, project context) -- and writes them itself with edit_file. In a non-interactive turn that write hits the ask permission tier and is refused. Correctly: the agent reports that accurately.
  2. The memory backend. Written by the host after the turn ends, extracted server-side. The agent never calls a tool for it and has no way to observe it.

"Remember X" triggers path 1. Path 1 fails, path 2 succeeds, and the agent reports on the only one it can see.

Two things make it worse:

  • Every turn ends with 1 turn(s) were not written to long-term memory because the memory service was unavailable. even when the episode is on disk. That sentence has contradicted disk truth in three separate acceptance runs.
  • The backend extracted the agent's own false report. The stored episode is titled Memory Save Failure for Green Brompton Bike 'Pickle' Noted on 2026-09-16 -- the content is right, the narrative is wrong, and it is now in long-term memory.

Why it matters

A user asking "did you remember that?" is told no by the component least able to answer. The recall works; the report does not. And the wrong report is itself remembered, so it can be recalled later as if it were a fact about the system.

Not a regression

None of this is introduced by the memory-backend seam work (#414); that branch touches neither path. It surfaced during acceptance for it, where the two paths could be compared against disk truth side by side.

Worth separating

  • whether the agent should be told about user_memory/ at all when a backend is configured
  • whether the end-of-turn sentence is reporting on something it can actually observe
  • whether an agent's claim about persistence should be extracted into memory as if it were an observation
Dominant language
Python
Stars
4.1k
Forks
94
Avg merge
9h 54m
Merged PRs (30d)
370

Getting set up

This project ships no dev container, Dockerfile or contributing guide, so setting up is up to you: start from its README, and see our first-contribution guide for the general steps.

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from EverMind-AI/Raven

All issues in EverMind-AI/Raven

Similar issues

More Python issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.