Give the agent persistent memory

Let eVi remember things between sessions — your preferences, project facts, decisions — and pick up durable, per-project context automatically. No custom code: the agent manages memory through its own tools, and project context is a Markdown file you check into the repo.

Uses: configuration (~/.evi/config.toml) + CLI · Applies to: the CLI, web, and desktop — they share one memory store.

Two mechanisms

Memory — durable notes stored as one Markdown file per topic under ~/.evi/memory/ (%USERPROFILE%\.evi\memory\ on Windows). Survives restarts. The model decides what's worth keeping and writes it itself.
Project context — an EVI.md (or AGENTS.md) checked into a repo, loaded into every system prompt while you work in that tree.

Memory: how the agent remembers

Memory is attached whenever the memory tool toggle is on (the default). The model drives it through five tools — you never call them, you just talk ("remember that I prefer tabs over spaces", "what do you know about project-x?"):

Tool	What it does
`remember(name, content, tags="")`	Save/overwrite a note (same `name` overwrites).
`recall(name)`	Return a note's full body.
`forget(name)`	Soft-delete — moves it to `~/.evi/memory/.attic/`.
`list_memories()`	JSON list of `{name, summary, tags}`.
`recall_by_tag(tag)`	JSON list of notes carrying a tag.

Each note is <name>.md; an auto-maintained INDEX.md is folded into the system prompt as a ## Memory index block, so the model always knows what is stored and calls recall(name) to pull a full body on demand. Names are [A-Za-z0-9_\-] (max 64 chars) and bodies cap at 64 KB. Because it's plain Markdown you can add or edit files by hand — /reload re-reads them mid-session.

Memory is on by default. To confirm the toggle, and to keep writes from prompting every turn:

# ~/.evi/config.toml
[tools]
memory = true   # attaches the memory store + tools to the agent

[auto]
auto_approve = ["fs", "code", "memory", "skills", "image"]   # memory = no prompt

Project context: EVI.md / AGENTS.md

Scaffold a project-context file in the current directory:

evi init                 # creates AGENTS.md (the cross-tool standard)
evi init --name EVI.md   # or eVi's own name; EVI.md wins if both exist

eVi walks up from your cwd and loads the nearest file, appending it to every system prompt — so you get per-project behavior just by cd-ing in. In a monorepo a root file plus a package-level one are layered (outermost first). Keep it to conventions, where things live, glossary, and hard "don'ts"; it's capped at 64 KB. See examples/EVI.md for a starter.

Curate memory automatically

After a stretch of work, have eVi review recent transcripts and prune or update long-term memory for you:

evi dream --hours 24     # review the last 24h and curate memory

This needs transcripts enabled ([tools] transcripts = true, the default). Deletions are soft — recoverable from ~/.evi/memory/.attic/.

The full field reference (compaction, per-user isolation, size limits) is in docs/features/memory-context.md.

Where to go next

Build an agent with the SDK — build_agent(..., enable_memory=True, enable_project=True) wires both mechanisms into your own agent.
Guard an agent's inputs and outputs — add safety layers on top.