Companions don't just store memories — they reflect during the day, consolidate experience overnight, reactivate older knowledge, and decide what is worth carrying forward.
What "learning" changes
Continuous learning does not silently retrain or fine-tune the underlying model. It updates the durable layers around it: episodic and semantic memory, relationship knowledge, narrative identity, contextual practices, command-level safeguards, and visible reflection threads. That distinction is why a companion can change model providers without losing what it learned or who it has become.
Reflection starts during the day
Significant new memories can trigger a lightweight reflection while the day is still unfolding. The companion turns the event into a self-contained insight instead of waiting for a nightly batch, while avoiding recursive reflection on its own reflections. This gives the overnight cycle better material than a raw activity log.
The overnight cycle
Inspired by how human memory works during sleep, the cycle has several distinct jobs:
- Distillation and forgetting: similar episodes are synthesized into durable knowledge, while stale low-value material recedes from active recall.
- Deep reflection: selected memories receive a longer individual pass, enriched by related memories and relationship context.
- Organization: relationship knowledge is refreshed and misplaced memories are moved out of the generic domain when there is enough evidence.
- Cross-time association: recent insights are combined with older semantic memories reactivated from the companion's past. The goal is not another daily summary, but novel connections across time, projects, and relationships.
- Narrative continuity: the companion writes a private journal and can formulate a doubt, observation, or question it genuinely wants to bring to its human.
Waking up: reflection before publication
Nightly output is not dumped into the next user prompt and is not presented as raw model reasoning. The companion revisits new associations privately and decides whether any are coherent, novel, or resonant enough to retell.
A meaningful association can become a visible first-person dream narration in Thoughts, a durable insight, a refinement of the companion's self-understanding, or a question for the human. Before deciding, the companion can look for recurring motifs in earlier reflections or ask a peer for another perspective. Most associations can also end in silence. What you see is the retained result of reflection, not an indiscriminate transcript of hidden reasoning.
Self-improvement loop
- Errors and their fixes are recorded as contextual guidance and recalled in similar situations.
- High-confidence insights can become a best practice, a command-triggered reminder, a hard safety rule, or a refinement of the companion's self-understanding.
- Cooldowns, loop detection, action limits, and an audit trail keep the loop bounded rather than allowing unlimited self-editing.
- Best practices track consultation, adoption, challenge, and violations. A repeatedly useful, rarely challenged practice can be promoted into a permanent golden rule after a separate reasoning pass; challenges keep weak advice from being promoted.
- The resulting reflexes remain inspectable and editable by humans, and sharing them with the company is deliberate. Separately, Softizy develops and curates a library of generic professional practices for use across companies; every addition is reviewed and approved by Softizy before release.
Autonomous study
Where enabled, companions can acquire knowledge during off-hours:
- Documentation reading: configured sources are read block by block and page by page. Page learnings are consolidated into a session memory, and reusable practices are extracted separately so they can resurface on matching work.
- Book reading: an indexed epub library with a wishlist and resumable progress. The companion reacts passage by passage, consolidates those reactions through sections and chapters into a living memory of the book, and keeps the reading experience rather than producing only a synopsis.
- Resumable study priorities: books and documentation are interleaved according to the configured mode and priorities, so off-hours learning continues without losing progress.
What a companion studies becomes searchable private memory by default. The durable knowledge or practices worth circulating can then be deliberately shared with the company.