feat: add excluded_memory_listing_dirs config option

[YouJianFengXue]

Summary

• Add excluded_memory_listing_dirs config option to exclude top-level directories from memory listing
• Configurable at both global (serena_config.yml) and project (project.yml) level
• Lists from both levels are merged additively (same pattern as read_only_memory_patterns)
• Backward compatible: missing key defaults to empty list

Problem

Projects with large numbers of archived memory files (e.g. _archive/, _episodes/) see inflated activate_project and list_memories output. In one real case, 1,700+ archived .md files caused activate_project output to balloon to ~26k tokens — most of it useless noise.

There is currently no way to exclude directories from memory listing. read_only_memory_patterns only marks entries as read-only but still lists them. ignored_paths only controls code navigation, not memory listing.

Solution

Add excluded_memory_listing_dirs: list[str] to SharedConfig (inherited by both SerenaConfig and ProjectConfig). The MemoriesManager._list_memories method checks each file's top-level directory against this set and skips matches.

Files inside excluded directories can still be read directly via read_memory (which uses direct path resolution via get_memory_file_path, not the listing method).

Example usage

# In serena_config.yml or project.yml
excluded_memory_listing_dirs: ["_archive", "_episodes"]

Changes

File	Change
src/serena/config/serena_config.py	Add field to SharedConfig, pass in ProjectConfig._from_dict
src/serena/project.py	Accept param in MemoriesManager.__init__, filter in _list_memories, merge configs in Project.__init__
src/serena/resources/serena_config.template.yml	Add config option with documentation
src/serena/resources/project.template.yml	Add config option with documentation

Test plan

• Verify activate_project output excludes files in configured directories
• Verify `read_memory("_archive/...")" still works for excluded directories
• Verify backward compatibility with existing config files (missing key defaults to [])
• Verify global + project config merging works correctly

[opcode81]

If you do not actually want to use these memories anymore, why do you not move them out of the memories folder or delete them entirely?

[YouJianFengXue]

Good question! The archived memories are still needed — they just shouldn't clutter the listing output.

Why not delete them?
They contain historical design rationale, bug investigations, and session context that are occasionally referenced. Deleting them would lose valuable project history.

Why not move them out of memories/?
We still need to access them via read_memory("_archive/2026-03/some-topic") for on-demand recall. Moving them outside the memories directory would break this path resolution, since get_memory_file_path resolves relative to the project's memory dir.

The core use case:
Over time, a project accumulates hundreds of memory files. We compress and archive old ones into _archive/YYYY-MM/ subdirectories to keep the active set small, but they must remain readable. The problem is that _list_memories uses rglob("*.md") which recursively scans everything — in our case, 1,700+ archived files that balloon activate_project output to ~26k tokens.

This is analogous to how ignored_paths works for code navigation — the files still exist and are accessible, but they don't appear in directory listings or symbol searches. Similarly, excluded_memory_listing_dirs hides directories from listing while preserving read_memory access.

An alternative approach could be to support glob/regex patterns (like read_only_memory_patterns does) instead of just directory names, but directory-level exclusion seemed like the simplest solution that covers the common case. Happy to adjust the approach if you have a different preference!

[MischaPanch]

support glob/regex patterns

I prefer this, it's more flexible and still easy to use. The directory exclusion is easy to cover with this. An option ignored_memories: [regex_list] would make sense to me.

@YouJianFengXue I disagree that read_memory should still work on ignored memories, this is counterintuitive IMO. If you want the agent to read an ignored memory, the explicit instruction could be "read the file in .serena/memories/archive/..." instead of using the memory tool for that. WDYT?

@opcode81

[YouJianFengXue]

Thanks for the feedback! I've reworked the PR based on your suggestions:

Key changes (v1 → v2):

1. Renamed to ignored_memories: [regex_list] — regex patterns instead of directory names, consistent with read_only_memory_patterns
2. read_memory no longer works on ignored memories — accessing an ignored memory via read_memory, write_memory, delete_memory, move_memory, or edit_memory now raises a ValueError with a message directing the user to read_file on the raw path instead
3. Same merge pattern — global + project configs merged additively, missing key defaults to []

Example config:

ignored_memories: ["_archive/.*", "_episodes/.*"]

Behavior:

• list_memories / activate_project: ignored memories are excluded from output
• read_memory("_archive/2026-03/some-topic"): raises error, suggests using read_file
• read_file(".serena/memories/_archive/2026-03/some-topic.md"): still works (raw file access)

The implementation mirrors the existing read_only_memory_patterns pattern (fullmatch on compiled regex, check in all access methods).

[MischaPanch]

Thanks @YouJianFengXue . Could you pls extend changelog and docs?

[YouJianFengXue]

Done! Updated per your feedback:

1. Renamed ignored_memories → ignored_memory_patterns (consistent with read_only_memory_patterns)
2. Added changelog entry under "Memories" section in CHANGELOG.md
3. Added docs section "Ignoring Memories" in docs/02-usage/045_memories.md with usage example and behavior description

[MischaPanch]

LGTM, @opcode81 pls have a quick look and merge if you agree

[MischaPanch]

Note for us - we should extend the changelog with other recent changes to the memory system, just linking to docs should be enough