Skip to content

ROADMAP.md

Latest commit

 

History

History
61 lines (51 loc) · 3.05 KB

ROADMAP.md

File metadata and controls

61 lines (51 loc) · 3.05 KB
Raw

Roadmap — what's next after 2.0

2.0 shipped the social layer: the willingness model (ambient chime-ins), human-paced multi-message replies, per-person impressions, and per-channel expression learning. This file tracks what's designed but not yet built, roughly in order of value-per-complexity on a 256 MB board.

Near — cheap, flat-file, high visible payoff

  • Mood as a decaying one-liner. A per-channel free-text mood ("feeling smug about that eval bet") updated with small probability per message, regressing to calm after a few minutes of silence, injected into the prompt. ~60 lines.

  • Emoji reactions as a social action. Let ambient turns choose to just react (🔥/💀/custom server emoji by name) instead of writing a message — the cheapest possible "she saw it" signal.

  • Time-based talk value. Chattier evenings, silent nights (board-local time). Trivial now that TalkValue exists.

  • Interest-driven focus. When a channel's willingness saturates repeatedly (a hot conversation she keeps joining), temporarily raise her presence there — shorter backoff, no coin flip for a few minutes — then cool down.

  • Holodex backends. The demo sandbox (Holodex) supports both static demos and verified container applications with no runtime network egress and hard CPU/RAM/process caps.

Mid — a day or two each

  • Concept-graph memory. Replace flat MEMORY.md retrieval with a tiny graph: nodes = topics with attached memory snippets, edges = co-occurrence with integer strength; built by a nightly model pass over sampled chatter, retrieved by keyword spreading-activation; forgetting = untouched edges lose strength. One gob file, no embeddings (token overlap instead). Also upgrades the ambient interest signal for free.
  • Conversation summaries (mid-term memory). Per-channel rolling digests written when a burst of chatter closes, so "what were we arguing about this morning?" works without replaying 80 raw messages.
  • Daily schedule. One model call each morning writes Vela's "day" (what she's tinkering with); referenced in prompts so quiet hours and current obsessions feel consistent.
  • Night consolidation. The board is idle at night — a 3 AM pass that merges near-duplicate memories, decays stale impressions, and prunes expression files.

Far — heavy infra the feel doesn't need

  • Vector/embedding retrieval — needs an embedding model and index; token overlap covers a private server fine.
  • Full knowledge-graph RAG over ingested documents.
  • Subprocess plugin runtime with RPC isolation.
  • Auto-collecting and describing the server's stickers/memes for use in replies — fun, but each sticker costs a vision call; revisit if a free local VLM lands on RISC-V.

Jargon dictionary (undecided)

Expression learning captures tone but not vocabulary semantics; a tiny jargon.json (term → meaning, learned the same nightly way) might be worth it once expression learning proves itself. Same guardrails: statistical, never from one occurrence, no slurs.