1) Basics: what “learning” means (2 modes)
A) In-session learning
I can use what you share in this chat to improve answers right now.
B) Cross-session memory
If you explicitly ask me to remember something (and it’s stable/useful), I can store it for future chats.
2) What is a “knowledge object”?
A knowledge object is a structured bundle of facts plus metadata so it can be reused consistently.
- Type: preference / profile / project / decision / definition / dataset / procedure
- Content: the actual facts
- Scope: this chat only vs. remember
- Source: you, a doc, a link, a meeting note
- Validity: time-bound or “until changed”
- Confidence: high/medium/low if uncertain
3) Templates (best way to ask)
Use these templates when you want me to treat input as structured knowledge.
Template 1 — Store as memory (persists)
Please remember this as a knowledge object:
Type: Preference | Profile | Project | Decision | Definition | Procedure | Other
Content: …
Scope: Remember for future chats
Source: User
Valid until: (optional date) or “until I change it”
Confidence: High | Medium | Low
Notes/Constraints: (optional)
Template 2 — Use now only (do not store)
Use this as a knowledge object for this conversation only (do not remember):
Type: …
Content: …
Scope: This chat only
Source: …
Valid until: …
Confidence: High | Medium | Low
Template 3 — Versioned object (recommended for projects)
Please remember this as a knowledge object:
ID: PROJECT-___-001
Version: 1.0
Type: Project | Decision | Procedure
Content:
- Goal: …
- Definitions: …
- Inputs: …
- Outputs: …
- Acceptance criteria: …
Scope: Remember for future chats
Source: User
Valid until: …
Change log:
- v1.0 (YYYY-MM-DD): created
4) Advanced uses of knowledge objects
4.1 Layered knowledge (base → overrides)
Use a “base” object for general rules and an “override” object for a specific task/client.
Knowledge Object (Base)
Type: Preference
Content: For market questions, start with a 3-line summary then a table.
Scope: Remember
Knowledge Object (Override)
Type: Preference
Content: For VLCC WS, include TD3C/TD15/TD22 and cite sources.
Scope: This chat only
4.2 Schema-driven objects
Define a schema once, then create multiple objects that conform to it.
Schema: DividendEvent
Fields: Ticker, Market, AmountPerShare, Currency, ExDate, RecordDate, PayDate, Source, AsOf
Knowledge Object
Type: DividendEvent
Content:
Ticker: VAR
Market: Oslo Børs
AmountPerShare: 1.209
Currency: NOK
ExDate: 2026-02-03
PayDate: 2026-02-12
Source: Company release
Scope: Remember
4.3 Decision log (auditability)
Keep a running record of decisions with rationale, date, and owner.
ID: DECISION-CHARTER-014
Version: 1.0
Type: Decision
Content:
- Decision: Use Baltic weekly TD3C as primary source; supplement with CME settlements.
- Rationale: Widely referenced benchmark; easy to cite.
- Owner: Tore
- Date: 2026-04-17
Scope: Remember
Change log:
- v1.0: created
4.4 Constraints & guardrails
Encode constraints to reduce rework (units, timezone, allowed sources, formatting).
Type: Guardrails
Content:
- Timezone: Europe/Oslo
- Units: NOK, USD/day, WS points
- Output format: Summary + table + sources
- Ask if missing: route, date range, market
Scope: Remember
4.5 Retrieval cues (make it easy to recall)
Add keywords so you can reference it quickly later.
Type: Project
Keywords: VLCC, TD3C, Baltic, WS, freight
Content: …
Scope: Remember
4.6 Expiring knowledge (time boxed)
Set an expiry so stale assumptions don’t live forever.
Type: Assumption
Content: Treat April 2026 VLCC market as volatile due to geopolitical risk.
Valid until: 2026-05-31
Scope: Remember
5) How to make a knowledge object (step-by-step)
- Name the purpose: what should this enable? (e.g., “consistent dividend answers”).
- Choose a type: Preference / Project / Definition / Dataset / Procedure.
- Define scope: remember vs. this chat only.
- Write content as bullet facts: short, unambiguous lines.
- Add metadata: ID, version, source, validity, timezone, units.
- Add acceptance criteria: how you’ll know it’s being used correctly.
Minimal “good” object
Please remember this as a knowledge object:
Type: Preference
Content:
- For dividend questions, always provide: amount/share, ex-date, pay date, and source.
Scope: Remember for future chats
Source: User
Valid until: until I change it
Confidence: High
Strong “best-practice” object (with schema + acceptance tests)
Please remember this as a knowledge object:
ID: SCHEMA-DividendEvent-001
Version: 1.0
Type: Definition
Content:
- Schema: DividendEvent
- Fields: Ticker, Market, AmountPerShare, Currency, ExDate, RecordDate, PayDate, Source, AsOf
- Output: Always return a one-line summary + a bullet list of key dates.
- Acceptance criteria:
1) If AmountPerShare is missing, ask.
2) If dates conflict, warn and ask which source to trust.
Scope: Remember
Source: User
Valid until: until changed
Change log:
- v1.0 (2026-04-17): created
6) How to test a knowledge object (checklist + test prompts)
Testing ensures the object is clear, non-contradictory, and produces repeatable outputs.
Checklist
- Clarity: could a stranger follow it without extra questions?
- Completeness: are required fields present (dates, units, scope)?
- Non-contradiction: does it conflict with other stored preferences?
- Scope correctness: did you specify remember vs. do not remember?
- Expiry: does it become stale? If yes, add Valid until.
- Acceptance criteria: did you define how success is measured?
Test prompts (copy/paste)
Use the stored knowledge objects and do a self-check:
1) List which knowledge objects you are applying.
2) Show the output.
3) Verify it meets the acceptance criteria.
If anything is missing, ask me exactly one clarifying question.
Validation test:
Given this knowledge object, generate two answers to the same question.
They should be consistent in structure and units.
Question: “What is the next dividend for VAR?”
Conflict test:
If two knowledge objects disagree, list the conflict and ask which one to prioritize.
Then provide the answer using the chosen priority.
7) Examples you can reuse
Example A — Shipping market preference
Please remember this as a knowledge object:
Type: Preference
Content:
- For shipping markets, reply with a 3-line summary first.
- Then include a table with routes, WS/levels, date, and source.
Scope: Remember
Source: User
Valid until: until I change it
Confidence: High
Example B — Dividend tracking procedure
Please remember this as a knowledge object:
ID: PROC-Dividend-Tracking-001
Version: 1.0
Type: Procedure
Content:
- When asked “next dividend”, do:
1) Identify listing/ticker.
2) Find latest company release.
3) Return AmountPerShare + ExDate + RecordDate + PayDate.
4) If not announced, state “not announced” and give next catalyst date.
- Output format: Summary line + bullets + sources.
Scope: Remember
Source: User
Valid until: end of 2026
Change log:
- v1.0 (2026-04-17): created
Example C — A full “DividendEvent” object
Use this as a knowledge object for this conversation only (do not remember):
Type: DividendEvent
Content:
Ticker: VAR
Market: Oslo Børs
AmountPerShare: 1.209
Currency: NOK
ExDate: 2026-02-03
RecordDate: 2026-02-04
PayDate: 2026-02-12
Source: Company release
AsOf: 2026-04-17
Scope: This chat only
Confidence: High
8) Quick Q&A prompts
Ask what I’m using
Before answering, list the knowledge objects you will apply and why (one line each). Then answer.
Force “no assumptions”
Answer using only what’s in the knowledge objects and what I provide. If something is missing, ask.
Forget an object
Please forget the stored knowledge object with ID: _____. Confirm it’s removed.
Update with a changelog
Update the stored knowledge object:
ID: _____
New Version: ____
Changes:
- …
Change log entry: (date + summary)
Then restate the updated object.