mikeyobrien Autoloop Memory
Loop memory is the durable store a loop reads into each iteration prompt. The default plugin is jsonl (.autoloop/memory.jsonl). A loop uses memory without forking core: pick a plugin with memory.kind in autoloops.toml.
Parse $ARGUMENTS to choose a verb. If none, list memory.
Verbs
# Persist
autoloop memory add learning "insight the next iteration should keep"
autoloop memory add preference Workflow "Always run tests before review.ready"
autoloop memory add meta smoke_iteration "2"
# Recall
autoloop memory list
autoloop memory find "pattern"
autoloop memory status
# Prompt-shaped render (same text the harness injects, no budget)
autoloop memory list
# Remove (tombstone; append-only)
autoloop memory remove <id>
autoloop inspect memory shows the same rendered block (--format md) or raw store (--format json).
Default jsonl store
Unless memory.kind is set, Autoloop uses the built-in jsonl plugin:
- Path:
AUTOLOOP_MEMORY_FILEorcore.memory_file(default.autoloop/memory.jsonl) - Entries: learning, preference, meta, tombstone
- Prompt injection is truncated to
memory.prompt_budget_chars(default 8000)
file is a second built-in plugin with the same jsonl file store.
To replace jsonl with Honcho, SuperMemory, Mnemosyne, or another store, install that product's Autoloop adapter and set memory.module. Autoloop does not ship those products.
[memory]
kind = "honcho"
module = "@example/autoloop-memory-honcho" # or "./memory/honcho.cjs"
prompt_budget_chars = 8000
The module exports createMemoryPlugin({ projectDir, kind }) (or default / plugin) implementing add / list / find / render. API keys stay in the adapter env. registerMemoryPlugin still works in-process.
When to write memory
- learning — a durable lesson discovered during a run
- preference — a categorized behavioral rule (always project-scoped)
- meta — a small key/value the next iteration should see
Do not invent a second memory product. Use these verbs against the configured plugin.