Session Replay for LLM Agents

Git for LLM agent sessions.

Upload a trace. Deltastore reconstructs the full session tree — every prompt, tool call, and decision point — with time-travel controls, branching, and shareable replay URLs. Storage costs a fraction of what naive logging needs, because Delta compresses it 91× behind the scenes.

Replay Step through every decision

Inspect prompts, tool calls, and intermediate state for any session.

Branch Fork any moment, explore counterfactuals

"What if the agent had chosen tool B here?" — branch and find out.

Share One URL, full context

Send a replay link. Teammates land inside the session at the exact state.

Compression 91×

Time-travel savepoints vs naive snapshots.

KV cache 10.6×

Single-snapshot compression, quality-bounded.

Serving-layer 32–113×

Effective storage savings at scale.

Setup < 60s

Upload JSONL → shareable replay URL.

Your agent broke in production. Can you reconstruct what happened?

Session state is ephemeral

The prompt, the tool outputs, the model's intermediate reasoning — most of it dies when the session ends. You're left with a chat log.

Naive logging is expensive

Full-fidelity traces of a long-running agent blow up storage budgets fast. Sampling loses the exact sessions you need to debug.

Counterfactuals live in your head

"What if the agent had picked tool B?" is a conversation today. It should be a fork you can actually run.

From JSONL trace to shareable replay in under a minute

1

Upload your trace

JSONL or OpenAI-style message logs. LangChain, LlamaIndex, AutoGen, CrewAI, or raw — we normalize it.

2

Deltastore reconstructs the tree

Every step becomes a node. Branches, retries, tool calls, and intermediate reasoning get first-class positions in the session graph.

3

Replay, branch, share

Scrub through time. Fork from any node to explore counterfactuals. Send a replay URL to your team — they land at the exact moment.

# minimal trace format
{"t": 0, "role": "system", "content": "You are a careful research assistant."}
{"t": 1, "role": "user", "content": "Find me papers on KV-cache quantization."}
{"t": 2, "role": "assistant", "tool": "web_search", "args": {"q": "KV cache quantization 2025"}}
{"t": 3, "role": "tool", "name": "web_search", "output": [...]}
{"t": 4, "role": "assistant", "content": "Here are three promising papers..."}

Why Deltastore storage costs less than naive logging

Deltastore sits on top of Delta — the continuity and compression engine underneath Solstice's production stack. The numbers are from the Delta benchmark harness, not marketing:

91× on time-travel savepoints

Traditional: 1,000 snapshots × 10 MB = 10 GB.
Delta: 1 base + 999 compact deltas = 110 MB.

10.6× on KV cache snapshots

Single-snapshot quality-bounded compression via eviction + quantization, full round-trip through transformer decoding.

32–113× at the serving layer

The number that matters once you're storing 90 days of replayable sessions across many agents.

What teams do with Deltastore

Debug production incidents

Replay the session where the agent hallucinated, find the exact prompt + tool output that flipped its decision.

Explain decisions to stakeholders

Send a replay URL to legal, ops, or the customer. They see the full context without you translating it.

A/B branch counterfactuals

Fork a session at the decision point, run the alternate tool, compare outcomes side-by-side.

Train evaluators

Capture the sessions that went right and the ones that went wrong. Use them as a gold-set for downstream evals.

Start free. Upgrade when you need more sessions, retention, or seats.

Free
$0/mo
  • 100 sessions / month
  • 7-day retention
  • Public replay URLs
  • 1 seat
Team
$299/mo
  • 10,000 sessions / month
  • 1-year retention
  • 5 seats, role-based access
  • Branch + fork workflows
  • Priority support
Enterprise
Custom
  • Unlimited sessions
  • Indefinite retention
  • SSO, audit log, on-prem
  • Policy enforcement + governed approvals
  • Starts with an Olympus Teardown
Book Olympus Teardown

Straight answers

What trace format do I upload?
JSONL where each line is one event: messages, tool calls, tool outputs, state changes. LangChain / LlamaIndex / AutoGen / CrewAI exports, OpenAI Chat Completions logs, or raw traces all work — we normalize them at ingest.
How is this different from Langfuse / Helicone / LangSmith?
Those are observability platforms — you watch your agents. Deltastore is a replay + time-travel environment — you step inside the session, branch from any moment, and run counterfactuals. The storage cost is also much lower because Delta compresses the session tree 91× vs naive snapshots.
Is Delta open source?
Parts of the engine (time-travel, savepoints, basic compression) will be open-sourced under Apache 2.0 as we stabilize the API. The compression kernel, KV-cache adapters, and multi-tenant hosting stay commercial.
What if I already have a session-tracing tool?
Deltastore imports from most of them. The replay experience is what differs — most tracing tools give you a flat timeline; Deltastore gives you a navigable session tree with branches and shareable state.
Data residency?
Hosted in US by default. EU region available on Team and above. Enterprise tier supports on-prem deployment inside your VPC — that path starts with an Olympus Teardown so we can scope your environment correctly.
Can I export my traces?
Yes, any tier. Full session tree dumps as JSONL + Parquet, with deltas reconstituted into whole snapshots. No lock-in.
Who's behind Deltastore?
Solstice AI Studio — the team running Delta in production under Solstice's own AI systems. Deltastore is the packaged version of that stack.
Get Started

Turn your next agent trace into a replay your team can step inside.

Free to start. No credit card. Upload your first trace and share the replay in under a minute.