Skip to content
DeepTutor EduHub

Memory

Memory is DeepTutor’s inspectable personalization system. It is deliberately not a hidden vector store — it is a file-backed, three-layer pipeline you can read, curate, and audit. Every high-level claim DeepTutor holds about you can be traced back to the raw events that produced it.

Where it is

Section titled “Where it is”

Open Memory in the left sidebar. The overview presents all three layers at a glance, with a live count for each, plus Refresh, Memory settings, and the entry into the Memory Graph.

The Memory overview: L1, L2, and L3 layer cards with counts, plus the Memory graph entry

The three layers

Section titled “The three layers”

Memory rises from raw events to synthesized knowledge. Each layer is a different fidelity of the same history.

LayerTagStorageRole
L1 · Workspace mirrorLIVEtrace/<surface>/<YYYY-MM-DD>.jsonlA snapshot of your live workspace plus the append-only raw event trace. Refresh records changes into L1.
L2 · Per-surface summariesCURATEDL2/<surface>.mdSurface-specific facts extracted by the consolidator, each citing the L1 events behind it.
L3 · Cross-surface synthesisSYNTHESISL3/<profile|recent|scope|preferences>.mdCross-surface propositions across a few slots — profile, recent timeline, knowledge scope, preferences — as hedged claims backed by L2 evidence.

Memory is tracked across the surfaces you actually use — chat, notebook, quiz, kb, book, partner, and cowriter. Because L2 cites L1 and L3 cites L2, nothing in your profile is unaccountable.

The Memory Graph

Section titled “The Memory Graph”

The graph view shows the whole pyramid at once: L3 synthesis at the centre, L2 facts in the middle ring, L1 raw traces on the outside. Hover any node for a preview; click to lock it and trace a claim from L3 → L2 → L1.

The Memory Graph: L3 at the centre, L2 in the middle, L1 traces on the outside, with a node preview

This is the tool to reach for when you want to answer: “Why does DeepTutor think it knows this about me?” — follow a synthesized claim down to the curated fact and then to the exact raw event.

Tuning the consolidator

Section titled “Tuning the consolidator”

Settings → Memory tunes the chunk-based consolidator: how many LLM rounds run per Update, Audit, and Dedup, how aggressively it chunks, and how strictly it validates references. Changes save to data/user/settings/main.yaml.

Settings → Memory: Update / Audit / Dedup budgets for the consolidator

ModeWhat it does
UpdateChunk-based incremental fact extraction. Set the L2 budget (per surface) and L3 budget (per slot). Per-doc overrides live in the workbench.
AuditLine-level edits checked against raw evidence — the integrity pass.
DedupIterative dedup over the full doc; stops early when an iteration emits zero edits.

CLI

Section titled “CLI”
Terminal window
deeptutor memory show L3
deeptutor memory show L2
deeptutor memory show profile
deeptutor memory show chat
deeptutor memory clear trace --force

Good practice

Section titled “Good practice”
  • Keep memory on for long-running learning workflows — that’s when it pays off.
  • Review L3 when DeepTutor’s tone or assumptions feel off; the graph shows you why.
  • Clear only the L1 trace for a surface if you want to drop raw history while keeping curated summaries.
  • In multi-user mode, each user gets scoped memory under their own workspace; a partner reads its own memory and its owner’s, but only writes its own.

See also

Section titled “See also”