nc://memory-select

Memory selection: keep the retrieved memories that fit the query and the token budget. A Nodecode rewrite of Pinutss/jev-memory-selector by Pinutss.

A retrieved memory list usually arrives too long: facts about other projects, another session's preferences, a note that expired last month. The original plugin sat between retrieval and the prompt and kept what was worth sending. This cell makes the same decision in the image, with no Python and no network at all.

From eval, the same call:

(memory-select:select "where does the app keep its data"
                      '((:id "db" :text "PostgreSQL, on the build host" :importance 0.7)
                        (:id "ui" :text "The settings page uses a dark theme")))

A memory answers with id and text (or content), and may carry created_at (ISO-8601), importance (0 to 1), tags, scope, expires_at and a retrieval similarity. A plist from eval reads the same way.

Get it

Tell the agent "add the memory-select cell", or run nodecode add nc://memory-select.

Set it up

Nothing is required: the defaults keep ten memories of about 4000 tokens, and every one of them is a member of the memory-select section in /settings or in config.jsonc:

{ "memory-select": { "max_results": 10, "max_tokens": 4000, "min_relevance": "0.15" } }

The three decimal members are text in the section, which carries no decimal type; a value that is not a number reads as its default. A member of the wrong shape refuses the start and says which one. Then (restart-cells "nodecode-memory-select").

Use it

Ask a question that leans on what was remembered — "which database did we pick, and where does it run?" — and an agent holding too many memories calls memory_select with the list and the question, then sends the selected ones.

How it decides:

• Relevance is lexical and deliberately simple: the share of the query's words (stopwords dropped, in English and French) that the memory's text and tags carry. A memory sharing no word with the query falls under min_relevance and is dropped as low_relevance.
• Freshness halves every recency_days; importance is the caller's own 0 to 1. The score is 0.5 × relevance + 0.3 × freshness + 0.2 × importance, and ties go to the newer memory, then the lower id, so the same list answers the same way every time.
• The budget is estimated at about four characters to the token. A memory that does not fit is dropped (budget) rather than cut in half, and a list past max_results is stopped (max_items).
• Duplicates: two memories that share dedupe of their words keep the first, and the second says duplicate_of:ID. A memory in another scope (out_of_scope), past its expires_at (expired), or carrying an id already seen (duplicate_input) is dropped before scoring.
• Contradictions are a warning, not a drop: MySQL against PostgreSQL, or SQLite against either, keeps both and marks the later one possible_conflict:ID.
• Keys that appear inside a memory answer as [REDACTED_API_KEY], [REDACTED_BEARER], [REDACTED_PASSWORD] or [REDACTED_PRIVATE_KEY], so a selection read out loud never repeats one.

What it does not do: it retrieves nothing (give it the candidates), it writes no prose, and it needs no key. (help :memory-select) answers the same guidance the model reads; memory-select::*primer* is the text, so a layer can reword it.