mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-11-16 19:04:08 +03:00
dfc2cbeaa1
agent/memory_manager.py mixes provider fan-out, background work and the text formatting that fences recalled memory for the model. The formatting half (sanitize_context, StreamingContextScrubber, build_memory_context_block and their regexes) is pure text work with no provider calls or profile state, so it moves unchanged to its own module. Callers import it from there; nothing in plugins/ imported it, so no re-export is kept. The only visible difference is the logger name on the "pre-wrapped context" warning (agent.memory_context). This leaves the manager room for the profile-ownership checks the next commit adds without crossing its line cap. Refs #134889 (PLG-11) Refs #140208
50 lines
2.4 KiB
Python
50 lines
2.4 KiB
Python
"""A recalled bullet is stated once per ``<memory-context>`` block.
|
|
|
|
Providers merge several stores, and the manager merges several providers, so one prefetch routinely
|
|
surfaces the same fact more than once. The composed block is stamped into the user row's
|
|
``api_content`` sidecar and replayed verbatim on every later request, so a byte-identical repeat is
|
|
paid once per turn for the life of the row — while telling the model nothing the block has not
|
|
already said.
|
|
|
|
Structure is not touched: headings, prose, blank lines and ``---`` rules survive as written, so a
|
|
provider's sections still read the way it wrote them.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
from agent.memory_context import build_memory_context_block
|
|
|
|
|
|
def _body(block: str) -> str:
|
|
"""The provider's own content, without the wrapper and the system note that precedes it."""
|
|
return block.split("]\n\n", 1)[1].rsplit("\n</memory-context>", 1)[0]
|
|
|
|
|
|
def test_a_repeated_bullet_is_kept_once_in_first_position():
|
|
"""Within a section a repeat is dropped; a section is delimited by a heading of ANY style.
|
|
RetainDB writes prose headings (``Profile:`` / ``Relevant memories:``); other providers write
|
|
markdown ones. In both shapes a placeholder bullet repeated under the second heading belongs to
|
|
that heading and must stay, or the heading is left claiming nothing."""
|
|
raw = "- alpha\n- beta\n- alpha\n- gamma\n"
|
|
|
|
body = _body(build_memory_context_block(raw))
|
|
|
|
# First-occurrence order, not a re-sort, and nothing else moved.
|
|
assert body.splitlines() == ["- alpha", "- beta", "- gamma"]
|
|
|
|
prose = "Profile:\n- None\nRelevant memories:\n- None\n"
|
|
markdown = "## A\n- (none recorded)\n\n## B\n- (none recorded)\n"
|
|
|
|
assert _body(build_memory_context_block(prose)) == prose
|
|
assert _body(build_memory_context_block(markdown)) == markdown
|
|
|
|
|
|
def test_a_bullet_with_continuation_lines_is_never_touched():
|
|
"""Two entries can share a headline and differ underneath it. Dropping one would re-parent its
|
|
provenance under the other and invent a record neither provider reported."""
|
|
raw = ("- prefers draft PRs\n (logged 12 Jan, source: supermemory)\n"
|
|
"- prefers draft PRs\n (logged 3 Feb, source: builtin)\n")
|
|
nested = "- Project A\n - status: active\n- Project B\n - status: active\n"
|
|
|
|
assert _body(build_memory_context_block(raw)) == raw
|
|
assert _body(build_memory_context_block(nested)) == nested
|