How Eldric knows things
Eldric keeps what it knows in several kinds of memory, each with its own rules for what goes in, who can read it and when it is forgotten. Searching your documents is one of them. This page shows all of them and links to the page that explains each in depth.
Facts and documents you give Eldric
Facts you tell Eldric ("my time zone is Vienna") are stored as you said them, each in a scope: private to you, a project, a group or the whole tenant. Documents go into knowledge bases, where the chat searches them and cites what it used. See Facts you tell Eldric and Your documents.
What happened in past tasks
When Eldric does a task with tools, it records an episode: which tools ran, how each step ended and what decided the outcome. You can say whether it worked. Episodes are found again only when someone asks, or before an agent task is planned. See Task episodes.
Ways of doing things
From checked episodes of the same kind of task, Eldric's learning cycle proposes methods by counting what keeps succeeding. A proposed method is used only after an administrator promotes it, and is then frozen. See How Eldric learns from episodes.
The conversation itself
Within a chat, the model sees the conversation so far. An administrator can also let Eldric add a few remembered facts to every message (proactive recall); that is off by default. Otherwise Eldric looks something up when it decides to, or when you ask.
Forgetting and compression
Recorded episodes are deleted after a retention period, 180 days by default. A forgetting policy, off until an administrator switches it on, compresses checked episodes first and deletes them later. See Memory: how it works.
Associative memory
Beside the exact store, Eldric keeps a compressed associative memory. Measured on 2026-10-29, it does not yet add anything to what recall returns: the facts you get come from the exact store. The numbers are on Memory: how it works, and what this memory is, beside Eldric's model files, on EMM and EXMM.
Giving Eldric knowledge
- A fact: tell the chat, or type
/remember <text>. - Documents: upload them to a knowledge base. For structured files, the knowledge-base wizard (
/api/v1/kb/wizard/*) reads the columns first. - Answers straight from documents: with
answer_from_memoryon, a knowledge base answers from its own documents without a language model; a model writes question and answer pairs once, when a document is added. - As an administrator: distil a knowledge source into memory (
/api/v1/distill/*).
How Eldric uses all this when it answers is on How Eldric reasons.