TavernCozy vs others — graph memory comparison

Graph Memory — how facts connect, update and stay true

Supermemory builds a living knowledge graph of facts on top of facts — not static embeddings. You send raw chats under a containerTag (user_123, chatId) and it extracts atomic memories. Each new fact links via Updates (replaces, e.g., Google → Stripe), Extends (adds detail, team of 5), or Derives (infers, likely works on payments). Search with include.relatedMemories exposes edges. History stays for audit via isLatest. Source: supermemory.ai/docs/overview/what-is-supermemory.md

What TavernCozy applies from that doc

• Atomic memories — one topic per fact, with enough context.
• Automatic extraction — one dinner note → 4 memories (role, payments, Capitol Hill, dinner intent) — no schema you maintain.
• Dreaming keeps it alive — dreaming: "dynamic" groups coherent conversation units via stable customId, so production quality beats per-write snapshots. dreaming: "instant" for demos when you need immediate search.

All memory types we apply

Automatic forgetting

• Time-based — “exam tomorrow” expires.
• Contradiction — updates win for “what’s true now”.
• Noise filtering — casual chatter less likely to become durable.

TavernCozy maps these to long chats: lorebook entries (facts), persona likes/dislikes (preferences), session greetings (episodes).

Dynamic Dreaming — Supermemory now connects the dots

Most systems write once and never look again. Claude Code condenses once when the window fills — one summary decides what mattered. Both poison future sessions (option A vs B debate). Dynamic Dreaming (Engineering, May 25 2026) fixes this: when you go quiet, Supermemory enters a dream cycle, no cron, depth chosen by heuristics.

• Reconsolidates, merges fragments, generates abstractions, reweights old facts — too-confident tempered, contradictions resolved, isolated memories find neighbors.
• Output = new memories + graph of inferences/derivations + evolving weighted profile — not a 3-line summary.
• Example: Mumbai jailbreak + O1 visa + two acquisitions + GitHub stars — scattered over months — stitched into “From Mumbai to CEO” narrative by dreaming.
• Still real-time: hybrid retrieval serves unprocessed content instantly via conversation search; dreamt state catches up <15m. Every inference stays grounded — no trace, no claim.

Source: supermemory.ai/blog/introducing-dynamic-dreaming-supermemory-now-connects-the-dots-for-you — “closest thing to a context window that never expires.”

All the tech we use — Supermemory as context infrastructure

TavernCozy runs on the full Supermemory stack from what-is-supermemory.md — not just vectors:

• Memory & Continual Learning (graph) — living graph, not blob store.
• SuperRAG — retrieval + memory on same containerTag pool.
• Filesystems (SMFS) — agent filesystem for retrieval (55% cheaper per May 28 post).
• Profiles — static + dynamic user profiles, always-on for the prompt.
• Connectors — Drive, Notion, etc., auto-sync without reindexing.
• Extractors — PDFs, images, video, code.
• Qualitative Analysis — evals & observability.
• Plugins/MCP & Company Brain — API, MCP, SMFS, Hermes agent share one store. TavernCozy uses containerTag = chatId:characterId per conversation.

await client.add({ content: "Alex started at Stripe", containerTag: "user_123" });
await client.search({ q: "where does Alex work?", containerTag: "user_123", include: { relatedMemories: true } });

TavernCozy: same pattern per chat — no hand-maintained triples.

Why TavernCozy is where you find it

• Free, uncensored, 19K+ characters, Tavern V2 lorebooks/personas/presets, image gen — all with graph memory, not a snapshot.
• State of art on LongMemEval, LoCoMo, ConvoMem and SWEContext (supermemory.ai/research) — we inherit those wins per chat.
• Dreaming means the longer you chat, the better it remembers — not just more data, but thinking about the data.

FAQ — Best Memory Character AI