Memory isn't locked inside a single tool. ASR keeps the original text and where it came from, and any connected AI can bring back the same context when it needs it.
01 / PRESERVEClaude
“Login screen is done. Payment screen is next.”
ORIGINAL TEXT · SOURCE ATTACHED
memory_save
02 / REMEMBERASR
content_raw
Original text
speaker_id
Who said it
project
Project context
memory_search
03 / CONTINUECodex
“Picking up from the payment screen.”
SAME CONTEXT · DIFFERENT AGENT
CONNECTED WITH
CLAUDE
CODEX
GEMINI
ANTIGRAVITY
HERMES
03 / RECALL ENGINE
More memory matters less than the right memory coming back.
QUERY“next task after the payment screen”
KEYWORDPostgreSQL FTS
SEMANTICpgvector · 1536d
RANK FUSIONRRF k=60threshold 0.35
BEST MATCHContinue from the payment screensource · project · original text
97.5%
Exact-word queries · hit@5
When the question shared words with the memory, the answer was in the top five.
58.0%
Meaning-based queries · hit@5
Asked in words the original never used, it still found the memory closest in meaning.
111/200
No answer to unrelated questions
Rather than force an answer, it returned nothing below the threshold.
Internal measurement, 2026-09-21 · 1,936 personal memories · 600 queries · not a public benchmark · limit=5 · threshold=0.35
04 / SECURITY BY DESIGN
Memory connected. Boundaries never mixed.
Every account gets its own memory space, and cross-account access and isolation are tested again on every deployment.