Repository navigation
PCS 0.9.180
PCS 0.9.180
This release brings together the changes since PCS 0.9.164: local speech that starts much faster on NVIDIA graphics cards, a longer pause before a spoken turn ends, coughs and room noise that are no longer typed into the conversation as words, web search that works again in conversations on Claude and ChatGPT, corrected instructions for the AI models, a Start summary that is no longer thrown away, experimental settings that let saved memories come to mind during a conversation (now also on OpenAI Realtime and Google Live, off unless you turn them on), memory jobs that no longer get thrown away while PCS Perspective is thinking, a clearer voice step in first-time setup, replies that are no longer lost when PCS saves a memory while it answers, a memory note that Claude no longer confuses with your own message, accepted memories that are no longer thrown away when only their summary is rejected, a Start summary that is no longer erased by a stretch of conversation with nothing to save, Vision that survives a single failed OBS capture, memory jobs that no longer get stuck on one batch for the rest of a conversation, a bug report that counts the tokens of each memory request, and a separate, cheaper model choice for Claude memory.
The release includes the Windows x64 application, ready to run offline with its pinned runtime and user guides, and the matching editable source. Tests, developer documents and build tools are only in the Source ZIP.
Memory that comes to mind (experimental)
Three settings change what the AI model is given from your saved memories. All three are off unless you turn them on, and Quick setup does not turn them on. Each takes effect at the next Start.
- Let memories come to mind on their own (Settings → Memory & sharing → Personal memory). PCS looks on your PC for saved memories and earlier turns related to what you just said.
- On Local AI, Meta, Claude, ChatGPT and OpenRouter, each time you type or speak it gives the model up to three, with no extra AI request. When nothing is clearly related, it gives nothing. On Meta the model is told to answer from such a note when it already answers the question, instead of searching again.
- On OpenAI Realtime (new), PCS checks the final transcript of what you said. A match that is ready in time goes with the first reply. If the model looks up a memory, the match can also go with the answer that follows. A match that arrives later can lead to an optional follow-up after the reply has played: after a question, or otherwise only when PCS Perspective may speak on its own. The model may speak or stay silent, and a follow-up can cost an additional paid Realtime reply. The model can also wait briefly for PCS's check, at most twice per turn, and each wait can cost another paid reply.
- On Google Live (new, paid personal-content connection only), PCS searches after each completed exchange and holds a match for up to two minutes. It is offered as you start speaking again, before PCS knows your new topic. If you move from a pickup plan to talking about the rain, Google may still receive the pickup detail; PCS asks it to leave the detail out on a new subject. Typing your next message clears the held match.
- In the feature's own spoken test on OpenAI Realtime, with synthetic speech and a made-up vault, 2 of 4 natural cases were answered correctly: one reply left out the saved detail and one contradicted it. That is why it stays off unless you turn it on.
- If the setting was already on in your vault, it now also applies on OpenAI Realtime and Google Live.
- Start with a fuller picture of what PCS knows (same place). At each Start PCS builds a short picture from your saved memories: your name, up to five ongoing subjects, and your last conversation. It replaces the three most recent memories. It is rebuilt every time, so a memory you corrected or deleted cannot stay in it.
- Show what PCS has in mind (Settings → Conversation → Session). The Memory stream panel lists what the model was given, by subject and size. It only shows; it saves and sends nothing.
- Both memory settings need "Learn from our conversations and recall personal memories".
- We first tried them on Local AI (Gemma 4 12B), in typed conversations, with a made-up test vault: three conversations of 14 turns, one each, judged by the developer and not blind. 13, 14 and 14 replies were judged good. Typed trials on cloud providers came later, in 0.9.169 (see below). The OpenAI Realtime and Google Live parts were tried only with synthetic speech in short scripted sessions, not by hand.
A separate model for Claude memory
- Claude memory writing, review and recall can use their own model: Settings → Connections → Memory. A new Claude setup with Recommended continuity uses Haiku 5.5 for memory and keeps Sonnet 5.5 for conversation. Existing setups keep what they have until you change it.
- This choice rests on one matched test with a made-up conversation: both models kept a correction, a suggestion attributed to PCS and a third-party estimate, and the estimated cost of memory writing and review was 92.8% lower with Haiku. It has not been tried on a real vault.
What changed for everyone
- The summary PCS starts a conversation with is no longer thrown away by unrelated changes, such as a PCS Perspective note, a reading bookmark or a settings save. It is still cleared when you correct, save or delete a memory, or change what PCS may learn or share.
- On Local AI with the default context, a memory lookup can now return twice as much (8,192 bytes, up from 4,096), so a lookup of two or three memories is delivered instead of withheld. A lookup that PCS started by itself no longer shows a notice.
- The instructions about adding to an existing notebook no longer contradict each other.
- A remembered-on-request note too large to keep no longer counts as a working memory check, so a memory-outage notice you have already seen is not shown again.
Memory jobs that finish
- In a 28-minute spoken conversation on Claude with PCS Perspective on, at least seven memory requests were thrown away part-way and 13 passages were still waiting to be learned at Stop. When PCS Perspective saved a step of its own thinking, any memory request already on its way was dropped and tried again later.
- PCS Perspective and memory formation now take turns. An unprompted Perspective step waits for a running memory job, and automatic memory formation waits for a running Perspective step. A pause to think that you asked for is not held back.
- On current Claude models the review that checks a proposed memory against its sources has more room for its answer (4,096 tokens, up from 2,048). One review of a long conversation had run out before giving a verdict.
- These repairs are covered by automated tests only. No conversation was run on this version.
Replies and memory notes on cloud providers
- We tried the memory settings in typed conversations on Claude, Meta, OpenRouter, Google and OpenAI, with a made-up test vault: one to three conversations of up to 14 turns on each, judged by an AI model and not blind.
- On Claude, the model twice took your own message for part of the memory note. The note now says where it ends and that your message follows. After this change, 14 of 14 Claude replies were judged good.
- When PCS saved a memory or a PCS Perspective thought while a reply was being written, the reply was dropped and the turn got no answer (twice in 14 turns on Meta). On Meta, Claude, ChatGPT, OpenRouter and Local AI a save now waits, for up to eight seconds, while a reply is being written. A correction, deletion or settings change you make yourself is not held back.
- In a rerun on Meta with an earlier form of this wait no reply was lost, and the released form lost no reply in one further Meta conversation. Three of that conversation's turns took 40 to 51 seconds because a memory lookup ran. No spoken conversation was tried with these changes, and the trials ran on the development copy, not on the released package.
Memories kept when only the summary fails
- A memory job proposes memories and a short summary for the next Start, and one review checks both. If the review accepted every memory but rejected the summary even after the job's one retry, the whole job was dropped and its passages waited for the next Start. In a recorded Claude conversation this happened once.
- Now the accepted memories are saved. The rejected summary is never stored. The summary you already had is kept if it does not mention a memory that changed; otherwise there is no summary until a later job writes one.
- This is covered by automated tests only. In one further recorded Claude conversation the case did not come up.
A Start summary that small talk no longer erases
- PCS starts each conversation with a short summary written by its memory jobs. A job that found nothing to save stored an empty summary over the old one, so a stretch of small talk or a goodbye erased it. After one 14-turn trial conversation the summary was empty on Claude and on OpenAI.
- Now such a job leaves the summary as it is. A job that writes a new summary still replaces the old one after review.
- In one further conversation each on Claude and OpenAI the summary was still there afterwards, though it described only the latest subject. Only the OpenAI conversation is known to have included a job with nothing to save.
Vision that survives one failed capture
- In a 22-minute spoken conversation on Claude, one OBS capture failed about 11 minutes 40 seconds in, and Vision stayed paused until the end. Why that capture failed is not known.
- When a capture fails in a way a fresh OBS connection can cure (OBS did not answer within five seconds, the connection dropped or was refused, or the request failed without a stated reason), PCS now connects again after two seconds and, if needed, once more after five. Vision pauses with the same notice as before only after the third failure in a row. A refused password, a missing scene, a wrong setting or an unusable image still pause Vision at once.
- The bug report now says how each capture failed (timeout, connection, refused password and so on), in fixed words only.
Memory jobs that no longer get stuck on one batch
- In the same conversation, a memory job on twelve waiting passages failed because the memory writer reached its output limit. The same twelve passages were then skipped on every later pass, 41 times, and nothing from the conversation was learned.
- When a draft is cut off at its limit (or, on OpenAI, Google and Local AI, is too large), PCS now tries again with half as many passages, and halves again if needed. A single passage that still does not fit is set aside so that newer conversation is learned. A cut-off draft is recognised on every provider and on Local AI.
- Any other failure that belongs to one batch (a refused, blocked, empty or incomplete draft, a request over a limit, or an unexpected error inside the job) now also sets that batch aside instead of blocking the queue until Stop. A set-aside batch is offered again after the next Start or by Process recent conversation. A failure of your account or the service (a refused key, quota or rate limits) still pauses memory processing as before.
- These repairs are covered by automated tests only. No conversation, OBS capture or paid request was run on the version that brought them (0.9.172).
Token counts for memory requests
- On Claude the memory writer still reached its output limit on batches of 12, then 6, then 3 waiting passages, and the bug report held no token counts that could show why.
- The bug report now lists the latest 16 memory requests: the output limit asked for, the input and output tokens the provider reported, the length of the visible reply, the time taken and how each request ended. Only counts and fixed words are recorded. A count a provider does not report (Local AI reports none) is left out, never shown as zero.
- On Claude the conversation footer now also shows "Last memory request total", as it already did on the other providers.
- This does not repair the output-limit failure. A later bug report on a real vault showed single-passage drafts of 11,798 to 15,381 visible characters using 4,748 to 5,918 of the 8,192 output tokens, so batches of three or more passages do not fit. This is not yet addressed.
First-time setup
- In the Voice step, the download buttons for the listening pack and the Kokoro voice need your vault, which setup creates at Finish. They used to be shown greyed out. They are now hidden until they can work, and the step says they appear right after Finish.
Faster Start with local speech
- Offline GPU listening no longer recompiles its graphics-card code on every Start. On the test PC (RTX 5080) listening and the Kokoro voice were ready together in about 6 seconds, against 32 seconds on 0.9.164.
- The first Start after installing still takes the longer time once.
- Local AI on NVIDIA cards also keeps its compiled graphics-card code between sessions, so the first picture of a session no longer waits about nine seconds. This applies to the recommended model and to a custom model.
- The compiled code is kept in
%TEMP%\PCS. It contains no audio, text or pictures.
Spoken turns
- On Meta, Claude, ChatGPT and OpenRouter with local speech, PCS now waits for 1.5 seconds of silence before ending your turn, as Local AI does (before: 0.8 seconds). The 60-second limit is unchanged.
- A cough, a click or room noise is dropped instead of being typed as "you" and answered. Spoken short replies such as "Thank you." and "What about you?" are kept.
- A short capture in which nothing was recognized now ends quietly, without a notice.
Web search and model instructions
- In a conversation on Claude or ChatGPT, every web search the model asked for came back as "The lookup delivery is not available." The search result now reaches the conversation, as it does on Meta and OpenRouter.
- This was checked with simulated Claude and ChatGPT connections, not with a live account. Pictures use the same search path and were not tried.
- Ten sentences in the instructions PCS gives its models were wrong or out of date. For example, they named a Materials button "Refresh list"; it is called Refresh documents. They were corrected or removed.
- We checked the text the models receive. We did not measure whether the models now answer differently.
Known limits
- The memory settings match words, not meaning. "My leg" does not find a saved knee injury; the model can still look it up.
- Even with a match, the model may overlook it, ask for another lookup or contradict it, as it did in one of four cases in the OpenAI Realtime test.
- On OpenAI Realtime a late memory or a wait for one can add paid requests. On Google Live a memory from the previous subject can be sent when you change subjects.
- When you ask about something PCS said and no word matches, PCS offers its most recent suggestion, promise or joke. In a long history that may be the wrong one.
- Memories longer than 900 characters do not come to mind on their own.
- The speech changes were measured with generated noise and synthesized speech, not with a microphone. A loud or speech-like noise can still be recognized as words.
- PCS still checks every local model and speech file in full on each Start. With a large local model this takes several seconds, longer the first time after Windows starts.
- If Windows clears its temporary files, the next Start compiles once more.
- Some instruction issues were found but left as they are. When you ask PCS to think, the instructions still say an empty reply is a good answer. Help still mentions Activity waiting.
- A passage whose memory job failed still waits until the next Start, or until you choose Process recent conversation.
- If every batch fails in the same way, PCS keeps trying newer ones (each a paid request on a cloud provider, at most one every 30 seconds) until about 1,000 passages are set aside, and then pauses automatic learning. A provider failure PCS cannot classify still holds up the queue as before.
- Claude, ChatGPT, OpenRouter and Meta reject a draft that is too large instead of splitting the batch; that batch is set aside.
- A Perspective pause you ask for can still interrupt a memory job that is running.
- On Claude about half of the memory jobs in the trials were turned down by the review for lack of evidence; their passages wait for the next Start.
- When a provider refuses a request with HTTP 400, PCS does not show the reason.
- The known limits of 0.9.164 remain.
AI-generated memories and reviews remain fallible. PCS provides inspection, original sources and correction controls; important interpretations should still be checked.
Upgrading from 0.9.164
- Save your work, choose Quit PCS, and wait for shutdown.
- Back up the entire stopped data folder, including any custom data directory. Copy original Materials separately if stored elsewhere.
- Extract the Windows ZIP into a new writable folder. Follow the included Backup & restore guide to copy or restore your existing data without merging data folders.
- Open Start PCS.exe, unlock with your existing passphrase and check your records. Keep the previous installation and backup while checking the update.
- If you use local speech on an NVIDIA card, the first Start may take about half a minute longer than later ones.
First launch installs the bundled Python runtime and pinned dependencies offline. Optional local AI models, speech packs and inference runtimes remain separate downloads.
Original Materials files are not encrypted by the vault; protect their backups accordingly.
Some newly saved features raise the vault's minimum supported PCS version. Perspective requires 0.9.122; newer native-provider records and Local research require 0.9.126; the newer cloud search index requires 0.9.130; native Claude/ChatGPT memory-thread suggestions require 0.9.134; a saved OpenRouter connection requires 0.9.151; and OpenRouter helper settings and Practice assessments larger than the old 4 KiB limit require 0.9.152. 0.9.155 to 0.9.180 add no new requirement. Opening the vault alone does not activate these requirements. The included vault compatibility guide lists the details.
Support PCS
If PCS is useful to you, you can help support its development. Support is optional.