Clarit.ai HomeReal memory, built inside the AI itself — not a database pretending to be memory, not a cache trick, not a training workaround. Your AI picks up where it left off, instantly, every time.
When you send a message, the AI re-reads your entire conversation from the beginning before it can reply. Every time. A 50-turn thread means 50 full re-reads. At scale, this is most of your AI bill.
The newest generation of AI models already compress the entire conversation into a small, fixed-size internal state. It's the same size whether the chat is 100 words or 100,000 words. Engram saves that state and restores it instantly — so the AI picks up exactly where it left off, without rereading anything.
Think of a context window as what the model reads. A snapshot is what the model is after reading it — a small file that captures the AI's current understanding, frozen in place. It stays the same size whether the conversation is five minutes or five months old. Save it, move it, restore it, and the AI picks up right where it left off without reading anything over again.
Not a copy of your conversation text. Not a screenshot. Not a log file. Saving the transcript just saves words — restoring one means the model still has to re-read them all.
Not a modified model. Not fine-tuning. Not new weights. The base model never changes — a snapshot is inference-time state, persisted between requests.
The compressed internal representation after the model processes the conversation — frozen and filed. Same fixed size at turn 1 or turn 1,000,000. Restore it and the model resumes exactly where it left off.
Measured on IBM Granite 4.0-H-Tiny (4B) · H200
Seven models across four vendors — IBM, NVIDIA, Alibaba, Mistral. Dense Mamba2 hybrids, MoE hybrids, LatentMoE, Gated Linear Attention, and pure Mamba2. Two generations of Qwen3 six months apart. All pass Engram's integration tests end-to-end.
| Model | Vendor | Architecture | Snapshot | Status |
|---|---|---|---|---|
| Granite 4.0-H-Tiny (4B) | IBM | Dense Mamba2 hybrid | ~56 MB | PASS |
| Granite 4.0-H-Small (32B) | IBM | Dense Mamba2 hybrid | ~146 MB | PASS |
| Nemotron-Cascade-2 (30B) | NVIDIA | MoE Mamba2 hybrid | ~47 MB | PASS |
| Nemotron-3-Super (120B FP8) | NVIDIA | LatentMoE hybrid | ~5.3 GB | PASS |
| Qwen3-Next (80B) | Alibaba | GLA + MoE · earlier generation | — | PASS |
| Qwen3-Coder-Next (80B) | Alibaba | GLA + MoE · evolved architecture | ~23.7 GB | PASS |
| Codestral Mamba (7B) | Mistral | Pure Mamba2 | — | PASS |
All snapshot sizes stay constant regardless of conversation length — a 1M-token conversation saves the same bytes as a 1K-token one. Codestral Mamba passes on Engram today; its snapshot figure publishes once SGLang merges the pure-Mamba2 loading fix Clarit contributed upstream (PR #22327).
Save and restore complete Mamba hidden state (conv_state + ssm_state tensors) via REST API. Constant-size snapshots regardless of context length.
VRAM → host RAM → disk with automatic promotion and eviction. Hot sessions restore in ~2 ms, warm in ~10 ms, cold in ~50 ms.
Capture state immediately after request completion but before memory is released. Previous approaches lost the state at cleanup.
REST and WebSocket APIs for checkpoint, branch, and restore. Integrates with OpenAI-compatible tooling and Vercel AI SDK.
Per-request correlation with futures-based system. 271 concurrent requests, zero cross-contamination, copy-on-write semantics.
Validated across dense Mamba2, MoE hybrids, LatentMoE, GLA, and pure Mamba2 — seven models across four vendors, including two Qwen3 generations six months apart.
Engram is open source. We work with teams and partners who care about AI that actually remembers — and we'd love to hear what you're building.