Long-running agents need a deliberate approach to context. The preserved IndusAGI TypeScript framework documentation describes context condensation, an event ledger, and branchable session storage inside the runtime. These are the basis of the memory behavior described here.
What the runtime manages
Before model calls, the runtime estimates the accumulated context. When the configured threshold is crossed, compaction summarizes an older part of the history while retaining a recent tail. The goal is to keep useful working context within the model's limits. Summaries can omit details, so important requirements and evidence should remain available for review.
The runtime also publishes run events and can persist a content-addressed session history as JSONL. Branchable sessions help a host resume work or explore a different continuation. See the runtime documentation and coding-agent session format.
What the memory facade does not provide
In this documentation snapshot, indusagi/memory is an empty facade. It does not export a standalone Memory class, VectorStore, Embedder, or embedding functions. The memory module overview explicitly distinguishes the facade from the real runtime machinery.
Do not infer that persistent sessions automatically provide semantic vector search or retrieval over an organization's documents. Retrieval is a separate design decision requiring suitable components, access controls, and source handling.
Configure and evaluate compaction
Review the edition's documented compaction settings. Test a conversation long enough to trigger condensation, check which recent turns remain verbatim, and verify that the agent can still recover the facts needed for its task. Save source references rather than relying solely on generated summaries.
Compaction changes token usage, but actual cost and quality depend on the provider, model, workload, thresholds, and summarization behavior. This guide does not claim a fixed percentage saving.
Choose the right building blocks
Start with the Framework overview, memory developer guide, and coding-agent compaction guide. Python and Rust have separate documentation; verify their current implementation rather than assuming feature parity.