| # WithIn Us AI — MemoryGenesis (God Level) 250K |
|
|
| Developer: **WithIn Us AI** |
|
|
| MemoryGenesis is a 250,000-example dataset designed to train LLMs to behave like **memory-first agents**: |
| - capture new information at runtime |
| - store durable and ephemeral memory safely (TTL) |
| - retrieve and cite relevant memories (RAG-style behavior) |
| - update, correct, merge, deduplicate, and compact memories |
| - resolve conflicts via provenance/recency/confidence |
| - enforce privacy (no secrets stored), redaction, and safe policies |
| - approximate "instant knowledge injection" **without weight updates** by using external memory tools |
|
|
| > Important: This dataset trains **runtime memory behaviors**. It does not perform real-time weight updates or replace model fine-tuning compute; instead it teaches a model to use external memory so new facts can be injected immediately. |
|
|
| ## Files |
| - `train.jsonl` — 245,000 examples |
| - `valid.jsonl` — 5,000 examples |
| - `sample_200.jsonl` — 200 examples for inspection |
|
|
| ## Record schema (JSONL) |
| Each line includes both instruct and chat formats: |
|
|
| - `prompt_instruct` / `response_instruct` (instruction SFT) |
| - `messages` (chat SFT: system/user/assistant) |
| - `task_type`: memory_write, recall, update, merge/dedup, compaction, TTL, privacy, schema, indexing, eval, Q&A |
| - `metadata.runtime_memory_only = true` and `metadata.no_weight_updates = true` |
| |
| ## Tool protocol (in-text) |
| Responses may include structured tool calls inside code blocks: |
| |
| - `memory.write` (key/value/tags/confidence/ttl_days) |
| - `memory.search` (query/k/tags) |
| - `memory.update` / `memory.delete` |
| - `memory.compact` |
|
|
| These are represented as JSON under `TOOL_CALL` or `TOOL_CALLS`. |
|
|
| ## Safety constraints |
| - Never store secrets (API keys, passwords, private keys). |
| - Store only safe derived information. |
| - Ask user to confirm when evidence is missing or conflicting. |
|
|
| ## Suggested fine-tuning usage |
| ### Instruct SFT |
| Map: |
| - input: `prompt_instruct` |
| - target: `response_instruct` |
|
|
| ### Chat SFT |
| Map: |
| - messages: `messages` |
|
|
| ## Evaluation suggestions |
| Measure: |
| - precision@k for retrieval |
| - staleness/conflict rate |
| - user correction rate |
| - privacy redaction compliance |
|
|
|
|
