Docs
Memory & Continuity

Memory & Continuity

How Viventium keeps facts, recent context, drafts, and project state useful without turning into stale prompt soup.

Why Memory Matters

Most chatbots only remember the current thread well enough to keep replying.

Viventium is aiming for something stronger: continuity that helps future work start sharper instead of starting from zero every time.

That means memory is not just "more chat history." It is a system for keeping the right kinds of context in the right place.

Continuity Layers

The source-of-truth docs point toward a memory model with several jobs, not one giant blob.

Durable facts

These are the stable things that should not be rewritten casually:

  • long-lived preferences
  • important facts about people, teams, and work
  • recurring constraints
  • established context that keeps showing up

Recent working context

This is the short-horizon layer:

  • what is active right now
  • what changed recently
  • what the assistant should still have in working memory

This is where recency matters the most.

Signals and patterns

This layer holds observations that may matter without pretending they are permanent truth.

Examples:

  • when you usually do your best work
  • recurring friction in a workflow
  • patterns that might deserve follow-up later

Drafts and in-progress work

Not everything should become a durable memory.

Drafts, partial ideas, and evolving plans should stay available without overwriting more stable knowledge.

Project state

Projects need their own continuity:

  • goal
  • context
  • constraints
  • current status
  • prior outputs
  • what should happen next

That is different from user memory, but it still belongs to the continuity story.

What Good Memory Should Feel Like

Good memory should make the system feel:

  • more prepared
  • more helpful
  • more consistent
  • less repetitive

It should not feel:

  • creepy
  • overconfident
  • stuffed with stale context
  • destructive when details change

Known Failure Modes

Memory is harder than it looks. These are the patterns Viventium is designed to avoid:

  • Rewrite-heavy updates — carelessly overwriting stable facts with temporary observations. Fix: prefer additive merges over destructive rewrites.
  • Recency drift — forgetting durable facts because recent context pushes them out. Fix: separate durable and working layers with different eviction rules.
  • Stale injection — stuffing old context into every prompt regardless of relevance. Fix: selective retrieval based on the current conversation, not a blanket dump.
  • Placeholder overwrites — partially parsed updates corrupting structured memory. Fix: validate payloads before writing.
  • Missing time context — memories without timestamps become ambiguous. Fix: inject canonical time context so the system always knows when "now" is.

Design Discipline

The principles that govern memory in Viventium:

  • Prefer additive updates over destructive rewrites
  • Keep canonical time context explicit in every memory operation
  • Separate durable facts from temporary notes by type, not by age
  • Use selective retrieval — inject what is relevant, not everything
  • Treat the source-of-truth config as the contract for what memory the system can access

Memory Architecture

Viventium supports two memory implementations:

v0.3 (semantic/episodic) — vector-based recall with explicit short-term/long-term separation. Stores snapshots, summaries, transcripts, and raw conversation data. Supports semantic search for relevant recall.

v0.4 (structured key-value) — key-value storage with structured prompt injection. Faster writes, simpler retrieval, better suited to the background agent model.

The important question is not only "where is memory stored?" but:

  • What gets remembered vs. what expires
  • What stays durable vs. what stays recent
  • What is actively injected vs. what is searchable on demand

That is why continuity in Viventium includes memory, drafts, schedules, and projects together — they are all part of the same continuity story.

Where This Shows Up In The Product

Continuity affects:

  • background follow-through after the first reply
  • morning briefings and scheduled updates
  • project work that unfolds across days or weeks
  • voice, chat, and messaging staying tied to the same context
  • connected workspaces becoming more useful over time

Keep Reading