Configuration

Configuration lives on two surfaces, deliberately separate.

Surface	Holds	Where
memsearch’s own config	Everything shared with the mesh: embedding provider and model, chunking, the compaction LLM	~/.memsearch/config.toml, with a project .memsearch.toml layered over it
pi-local environment	Only how pi drives memsearch	the PI_MEMSEARCH_* variables below

Anything that changes what the store looks like to another agent belongs upstream in memsearch’s config, so Claude Code, Codex, OpenClaw, OpenCode and pi keep reading one store the same way (ADR 0001, mesh parity). pi-memsearch adds no config of its own for those.

Variables

Variable	Default	Effect
PI_MEMSEARCH_CAPTURE	on	off disables automatic capture; memory_write keeps working
PI_MEMSEARCH_CAPTURE_MODEL	cheapest model of the session provider	Distillation model, as <id> or <provider>/<id>
PI_MEMSEARCH_SNAPSHOT	on	off disables snapshot injection
PI_MEMSEARCH_AUTO_CONTEXT	off	on enables per-prompt injection via a warm sidecar (~1 GB resident)
PI_MEMSEARCH_SEARCH_TIMEOUT_MS	30000	Per-attempt timeout for memory_search (each cross-repo invocation too)
PI_MEMSEARCH_COMPACT_TIMEOUT_MS	300000	Per-attempt timeout for memory_compact (LLM pass plus reindex)
PI_MEMSEARCH_SCAN_ROOTS	unset	:-separated directory roots scanned for other projects’ memory stores; required by cross-repo recall
PI_MEMSEARCH_STORE_CMD	unset	Command printing the store path (memory-dir), the collection (collection) and optionally the index-state dir (state-dir); outranks MEMSEARCH_DIR. Contract and reference
MEMSEARCH_DIR	unset	memsearch’s own scope override; the memory store and collection follow it. A relative path resolves where the memsearch children run

MEMSEARCH_DIR is the one entry that is not pi’s. It is memsearch’s own variable, listed here because pi mirrors its resolution order exactly rather than inventing a second one.

Auto-context

Off by default, and the only setting with a standing cost.

While it is on, a per-session sidecar holds the embedding model warm and takes 0.7 to 1.0 GB of resident memory. Every prompt gets a 300 ms hard cap. A deadline miss, an empty result or a locked store all degrade to no injection, so a prompt never waits on memory.

Remote embedding providers will often miss that cap. Three consecutive misses count as a crash, and past the respawn cap auto-context switches off for the rest of the session rather than spending 300 ms on every prompt.

Every tunable constant, and how the sidecar borrows the Milvus Lite lock for milliseconds at a time: the runtime page.