Using "Keyword" calls to save data

[ITFinesse]

Hello,
So as discussed, I added a dedicated instruction file section containing my own keywords and rules, which directs the memory executable on how to categorise information that I add (or that it captures automatically). Just allows me to input useful info as i go, and if the previous chat/prompt fixed something, tell it on the next prompt to remember it AND next steps, in one prompt.

In chat, I use keywords such as MOSRT ("remember this") and MOSRL ("remember last"). These indicate whether the information should be stored as:

• Workspace/Project-specific information, or
• Global information that is generally useful across projects.

Information is then categorised further, for example as:

• fixes
• rule
• howto
• pattern
• research
• other resource

Before creating a new memory entry, the system checks whether the information already exists:

• If a potential duplicate is found, it highlights the conflict and asks for confirmation.
• If an existing entry is found, it updates that record.
• If no entry exists, it creates a new one.

After processing, it displays a confirmation summary including:

• Created or updated status
• Scope (Project/Workspace or Global)
• Goal (if applicable)
• Category (for example, "Fix")
• A one-sentence description of the stored item

This provides a consistent structure for storing, updating, and retrieving knowledge while reducing duplicates and categorising information automatically.

The file:

• Check MemoryOS and search for specific relevant information, Read it before regression, bug-fix, test-planning, validation, or behaviour-changing work.

MemoryOS

Use only:

& "C:\memoryos\.venv\Scripts\memory.exe"

• Never use bare memory, .codex\memories, ad-hoc notes, or another memory system.
• Search before bugs, regressions, failures, deployment, security, performance, or non-obvious work.
• Save only for explicit MOSRT or MOSRL.
• Never save secrets, sensitive data, guesses, trivial edits, transcripts, or unverified claims.
• This workspace always uses --project <project name>; never use categories such as fixes as --project.
• Categories: fixes, rules, howtos, patterns, research, resources. Start --decision with one.

MOSRT / MOSRL

• Verify source files/facts, then search for an equivalent note.
• No equivalent: create one.
• Equivalent unchanged: do not save; report no update needed.
• Equivalent changed: show its title, ID, and path; use a documented update method, or ask before replacing it. Never create a duplicate.
• MOSRL must start the prompt and saves only the previous completed verified fix (Category: fix).
• Use explicit learning only:

& "C:\memoryos\.venv\Scripts\memory.exe" learn --project "<Workspace>" --goal "<subject>" --decision "Category: <category>. <concise verified decision>."

Report

Use exactly one:

-  Saved: created a MOSRT MemoryOS note.
-  Saved: updated a MOSRT MemoryOS note.
-  MemoryOS: no update needed.

• For updated/no-update: Existing record: Project <project>, Goal <goal>, Category <category>, Decision <one sentence>, Record <filename>.
• For saved: Project <project>, Goal <goal>, Category <category>, Decision <one sentence>, Record <filename>.
• Copy values exactly; do not invent abbreviations.
• Do not say saved/updated without Learned: <path>.

Example:

Prompt:
MOSRT, its our regression/bug fix work for this workspace/project RegressionChecks.md

Response:
I’ll quickly verify RegressionChecks.md exists, check available MemoryOS commands, then run a minimal search/update flow and only record a MOSRT memory note if there isn’t already an equivalent entry.

• I found equivalent MOSRT records already in MemoryOS. I’ll confirm the exact existing note metadata, then report that no new save is needed per your rule.

• MemoryOS: no update needed.

• Existing record: Project XYZ. Goal XYZ regression and bug-fix contract, Category rules, Decision RegressionChecks.md is the canonical regression/bug-fix contract and must be read before regression work/bug fixes/test planning/validation/behavior changes, Record 2026-07-28-task-learned-XYZ-regression-and-bug-fix-contract.md.

[vetrovk]

Thanks, this is a very useful example of an explicit, user-controlled memory workflow. MemoryOS already supports context, search, explicit learn --project --goal --decision, curator duplicate handling, and stable updates for PR and OSS candidate records.
For now, I think MOSRT and MOSRL are best kept as conventions in an individual agent workflow rather than added to the core as special keywords. The general learn command does not yet have a public generic update-by-id contract, and automatic text-based replacement would be risky.
A small documentation example for this workflow could be worthwhile. If you need a safe upsert flow for ordinary notes with an explicit identity, that would be a separate product discussion.