Context assembly engine
On every request, Trove chooses what goes in the window, not your team. Recency- and relevance-weighted, packed to your token budget for this agent, this task, this moment.
Your agent forgot everything.
Again.
Every new session starts from zero. Trove prepends what your agent already decided, token-budgeted, not the full prior session.
Works with
OpenAI
Anthropic
LangChain
CrewAI
Vercel AI SDK
Your agents already produce the memory.
It resets every time a session ends.
i.
Monday, retry-agent settles on full jitter and a 4-retry cap. Thursday, the same agent opens a blank window and re-researches the same tradeoffs.
ii.
What one agent learns disappears when the session ends. The next agent in your stack starts from zero, every time.
iii.
Most memory tools need an SDK install, schema design, and custom plumbing before useful context ever reaches a prompt.
iv.
When a response is wrong, there is no record of what reached the model at call time. You reconstruct the window by hand.
Cross-agent · token budget
Don't make implement-agent pay for architecture again.
Thursday's task is one line. Without Trove, implement-agent re-reads the PR thread or re-debates backoff strategy. Trove prepends 287 tokens from Monday's architecture-agent run, not thousands of tokens of thread paste.
Mon 2:11pm
architecture-agent
MonThu
Thu 9:47am
implement-agent
Task
Add retry logic to the payments service HTTP client.
Prepended from Monday · architecture-agent 287 tokens
Chose full jitter over exponential backoff. Max 4 retries, base delay 200ms. Do not retry on 4xx. Reference: AWS Builders Library retry guidance.
- Token budget
- Append-only log
- Your keys
- Provider-neutral
Not a database. A context engine.
Storage is a commodity. What matters is what context surfaces, and when.
Automatic entity extraction
Users, tickets, services, decisions parsed from every gateway call. No manual tagging.
Structured knowledge graph
Derived from the event log: typed relationships with timestamps. This is what Trove searches when it enriches the next call.
Cross-agent awareness
What one agent learns is available to the next. No shared-memory contracts or handoff glue.
Append-only event log
Raw record of every gateway call. Replay and audit exactly what was prepended when a decision shipped.
Provider-neutral
OpenAI today, Anthropic tomorrow. Switch models without losing state or history.
Every competitor is building a feature.
Trove is neutral infrastructure. The one position a model provider can't take.
| Trove | OpenAI / Anthropic | Vector DBs | LangChain / CrewAI | Memory SDKs | |
|---|---|---|---|---|---|
| Multi-agent, cross-session memory | ✓ | ✗ | ~ | ~ | ~ |
| Provider-neutral infrastructure | ✓ | ✗ | ✓ | ~ | ~ |
| One-line proxy integration | ✓ | ✗ | ✗ | ✗ | ✗ |
| Context assembly (not just storage) | ✓ | ~ | ✗ | ~ | ~ |
| Framework-agnostic | ✓ | ✗ | ✓ | ✗ | ~ |
| Stateful agents in under 5 minutes | ✓ | ✗ | ✗ | ✗ | ✗ |
One line to integrate.
Swap your client's base_url. No SDK, no schema. Trove holds the memory layer; your provider keys stay in your environment.
- Same OpenAI-compatible client. Trove sits between your agent and the provider.
- Every call is logged, matched memory is prepended, and the request is forwarded.
- Works with the OpenAI SDK, LangChain, CrewAI, and Vercel AI SDK.
Join the beta.
Leave your email and we'll send gateway access and setup details. One base_url swap, then memory carries across sessions.