Memory

Mister Morph memory uses an append-first WAL (Write-ahead logging) model plus asynchronous projection.

WAL

In plain terms, anything that happens is first written as raw data to a jsonl file in order. That is why WAL is the real source of truth.

The path is memory/log/, and file names follow the format since-YYYY-MM-DD-0001.jsonl.

When a WAL file reaches a certain size, it is rotated into a file ending in .jsonl.gz.

Projection

Mister Morph builds simple projections from WAL into two targets:

• memory/index.md (long-term memory)
• memory/YYYY-MM-DD/*.md (short-term memory)
  • Short-term memory files are isolated by channel. For example, memories from different Telegram group chats do not mix.

Projection means reading WAL for a period of time, summarizing it with an LLM, and writing the result into those target files.

The projection checkpoint file is memory/log/checkpoint.json, and it looks roughly like this:

{
  "file": "since-2026-02-28-0001.jsonl",
  "line": 18,
  "updated_at": "2026-02-28T06:30:12Z"
}

Injection

When the following config is enabled, part of the projected memory is injected into the prompt:

• memory.enabled = true
• memory.injection.enabled = true

memory.injection.max_items controls the maximum number of injected items.

Notes

1. If memory projections are damaged, rebuild them from WAL by deleting memory/log/checkpoint.json and letting the Agent continue running.
2. In production, keep file_state_dir on persistent storage so runtime state, including memory, survives.