AllSource vs Letta
Letta (formerly MemGPT) gives agents a stateful memory loop. AllSource gives that loop a durable, queryable foundation with full event history and point-in-time replay.
| Capability | AllSource | Letta |
|---|---|---|
Durable event store (survives restart) AllSource Core is the database: a Rust WAL with CRC32 checks, configurable fsync, and columnar Parquet persistence. Event data survives restarts under the selected durability mode. | WAL + Parquet | unknown |
Full event provenance / replay Every change is an immutable event you can replay to reconstruct any past state. | ||
Temporal / as_of queries Point-in-time projections are a first-class query, not a snapshot you have to manage yourself. | ||
Core indexed-read latency Published reference benchmark for Core indexed reads. It does not measure graph, vector, hybrid, or end-to-end agent recall. | 11.9μs (p99) | unknown |
Ingestion throughput Published Core batch-ingest reference. We do not publish competitor throughput figures we cannot verify. | 469K events/sec (reference batch benchmark) | unknown |
Offline / embedded mode Run AllSource in-process as a library — no separate server or network hop required. | ||
MCP tools for AI agents The default tenant connector exposes 55 tools. Fleet and administrative controls bring the exposed total to 73. | 55 tenant / 73 fleet | unknown |
x402 per-call agent payments Native x402 micropayments let agents pay per call instead of negotiating a seat-based contract. | ||
License / self-host Self-host Apache-2.0 Core and community components for free; enterprise components use BSL 1.1. Hosted tiers are optional. | Apache-2.0, self-host | unknown |
AllSource figures are from our own benchmarks and documentation. Cells marked “unknown” or “varies” are values we do not publish for Letta rather than estimate. Based on public documentation; corrections welcome.
Primary sources checked
Pick AllSource if…
- You want a durable store of record under your agents, not just an in-context memory manager
- You need point-in-time queries and full replay for audit or debugging
- You want a published microsecond-range Core indexed-read benchmark and the option to embed or self-host under Apache-2.0
- You want per-call (x402) economics for autonomous agents
Pick Letta if…
- You want an opinionated agent runtime with memory management built in
- You are standardizing on Letta's agent abstractions and tooling
- You do not need a separate durable event store of record
AllSource vs Letta: FAQ
Build on a durable event store
Start on the hosted Indie plan, or self-host the whole stack for free under Apache-2.0. Either way you keep full event provenance. Published microsecond-range results cover Core indexed reads, not end-to-end memory recall.