Advanced memory settings
Tuning options for memory, consolidation, the notebook location, and search.
Most people never open this page. Memory is on out of the box, and the two settings you are most likely to want — whether it runs at all, and which models it uses — are in Settings → Memory System in the app.
What follows is the tuning that has no UI. Put overrides in ~/.suzent/config/default.yaml
(machine-specific paths belong in ~/.suzent/config/local.yaml); both are read after the
shipped config/default.example.yaml, so anything you set there wins and survives an
update. Keys are case-insensitive — MEMORY_ENABLED and memory_enabled are the same
setting.
The basics
memory_enabled: true # Also the toggle in Settings → Memory System
markdown_memory_enabled: true # Write facts to /shared/memory/ as Markdown
extraction_model: gemini/gemini-2.5-flash # "" for heuristic extraction, no model call
user_id: default-userThe extraction model runs every few exchanges, so a small fast model is the right choice here. Leave it unset to use the default chat model.
These settings change how memory works. To change what Suzent remembers, tell it in a conversation or edit the notebook pages; see Managing what's remembered.
Consolidation
How often the organizer runs, and how much it takes on. See How memory is tidied.
memory_consolidation_enabled: true
memory_consolidation_interval_seconds: 1800 # How often to check whether a run is due
memory_consolidation_min_hours: 24 # Minimum wait between runs
memory_consolidation_min_facts: 20 # New facts needed before a run is worth it
memory_consolidation_min_confirmations: 25 # Repeated facts that justify a run on their own
memory_consolidation_max_days: 14 # Days of logs read in one run
memory_consolidation_max_retries: 3 # Attempts before a stuck batch is skipped
memory_consolidation_timeout_seconds: 600
memory_consolidation_model: null # Legacy; set the Memory organizer role in Settings → Model Roles instead
memory_map_max_lines: 30 # Size cap on the MEMORY.md mapHow facts are picked out
There is nothing to set here. Facts about you are saved only when they are in your own words, lessons only when the chat shows both the failure and the fix, and facts from files or pages only with their source.
The notebook audit runs on its own, slower schedule, and only once consolidation is caught up:
memory_lint_enabled: true
memory_lint_min_days: 7Where the notebook lives
notebook_dir: <data-dir>/notebookTo use an existing Obsidian vault instead, mount it over /mnt/notebook, replacing the
default entry. This takes precedence over notebook_dir, which keeps pointing at the
default path. Being machine-specific, it belongs in ~/.suzent/config/local.yaml:
sandbox_volumes:
- "C:/Users/you/Documents/MyVault:/mnt/notebook"Search
embedding_model: gemini/gemini-embedding-001
embedding_dimension: 3072 # Must match the model; 0 = detect from it
embedding_timeout: 30 # Seconds before a slow provider is given up on
lancedb_uri: <data-dir>/memoryPick the embedding model in Settings → Memory System — the list is filtered to what your configured providers actually offer, so it stays correct as you add or remove keys.
Changing the embedding model means re-embedding everything you have stored: one call per chunk across your whole memory. It is not a free switch, and the dimension has to change with it.
Sessions
session_daily_reset_hour: 0 # UTC hour for a daily reset (0 = off)
session_idle_timeout_minutes: 0 # Reset after inactivity (0 = off)Context window
max_context_tokens: 0 # 0 = use whatever the active model supportsLong conversations are compacted automatically once they reach 500k tokens or
80% of the model's context window, whichever comes first
(context_compaction_trigger_tokens, context_compaction_trigger). A 128k model
is compacted at about 102k and a 200k model at 160k. A 1M-token model is
compacted at 500k: past that, even the strongest models get noticeably worse at
recalling several details from a long conversation, and every message pays for
the whole history. If Suzent doesn't know a model's window, it assumes 200k
tokens.
Set max_context_tokens to a non-zero value to cap that budget, for example to
keep costs down. It can only lower the limit, never raise it.
The summary that replaces older messages is written by the model you are
chatting with, reusing the conversation it has already read, so most of it
comes from the provider's prompt cache and costs a fraction of the normal price.
Once the conversation reaches 350k tokens or half the window, whichever comes
first (context_precompaction_trigger_tokens, context_precompaction_trigger),
Suzent starts writing the next summary in the background. That happens during a
long run as well as when a reply finishes, so the summary is usually ready
before compaction is due and you don't wait for it. A prepared summary is kept
on disk, so a restart doesn't throw it away.
Providers only keep that cache for a while (Anthropic 5 minutes, OpenAI and Gemini about 5 minutes, DeepSeek hours). If you come back to a conversation after its cache has expired and it is still past that size, Suzent sends the prepared summary instead of the full history, since the full history would be billed at the normal price again. While the cache is still warm, nothing changes until compaction is due. Suzent also learns from your own usage: when a provider still serves the cache after a longer pause, it trusts that longer lifetime from then on.
When a summary has to be written on a cold cache, Suzent compares the price of the conversation's model against the memory extraction model and uses the cheaper one.
Seeing compaction
In the chat, a compaction shows as a thin divider at the point where it happened, with the context size before and after. A line under it says when the summary was prepared in advance, or that it was used because you were away long enough for the cache to expire. Show summary opens what the older messages were folded into. A failed compaction shows in red and leaves the history as it was.
The Context panel shows when the conversation was last compacted, and whether the next summary is being prepared or is ready.
Other settings:
anthropic_prompt_cache_ttl: 1hasks Anthropic to keep the cache for an hour instead of 5 minutes. Writing to it costs twice the input price instead of 1.25 times, so it pays off when you often pause between messages.context_precompaction_trigger: 0turns preparing summaries off.context_compaction_trigger_tokens: 0andcontext_precompaction_trigger_tokens: 0drop the token sizes and use only the share of the window.compaction_reuse_prompt_cache: falsehas the memory extraction model write every summary.
Rebuilding the search index
The Markdown files are the real memory; the search index is built from them. If
search starts returning odd results, open the Memory panel, view
MEMORY.md, and click Reindex. Nothing is lost, but a full rebuild
re-embeds every file, so it takes a while and costs embedding calls.